suno-ai
/
bark
mirror of https://github.com/suno-ai/bark.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

first commit

pull/4/head
Georg Kucsko 2023-04-09 13:21:02 -04:00
commit ea9a687004
18 changed files with 1769 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

400
LICENSE Normal file
View File

@ -0,0 +1,400 @@
Attribution-NonCommercial 4.0 International
=======================================================================
Creative Commons Corporation ("Creative Commons") is not a law firm and
does not provide legal services or legal advice. Distribution of
Creative Commons public licenses does not create a lawyer-client or
other relationship. Creative Commons makes its licenses and related
information available on an "as-is" basis. Creative Commons gives no
warranties regarding its licenses, any material licensed under their
terms and conditions, or any related information. Creative Commons
disclaims all liability for damages resulting from their use to the
fullest extent possible.
Using Creative Commons Public Licenses
Creative Commons public licenses provide a standard set of terms and
conditions that creators and other rights holders may use to share
original works of authorship and other material subject to copyright
and certain other rights specified in the public license below. The
following considerations are for informational purposes only, are not
exhaustive, and do not form part of our licenses.
Considerations for licensors: Our public licenses are
intended for use by those authorized to give the public
permission to use material in ways otherwise restricted by
copyright and certain other rights. Our licenses are
irrevocable. Licensors should read and understand the terms
and conditions of the license they choose before applying it.
Licensors should also secure all rights necessary before
applying our licenses so that the public can reuse the
material as expected. Licensors should clearly mark any
material not subject to the license. This includes other CC-
licensed material, or material used under an exception or
limitation to copyright. More considerations for licensors:
wiki.creativecommons.org/Considerations_for_licensors
Considerations for the public: By using one of our public
licenses, a licensor grants the public permission to use the
licensed material under specified terms and conditions. If
the licensor's permission is not necessary for any reason--for
example, because of any applicable exception or limitation to
copyright--then that use is not regulated by the license. Our
licenses grant only permissions under copyright and certain
other rights that a licensor has authority to grant. Use of
the licensed material may still be restricted for other
reasons, including because others have copyright or other
rights in the material. A licensor may make special requests,
such as asking that all changes be marked or described.
Although not required by our licenses, you are encouraged to
respect those requests where reasonable. More_considerations
for the public:
wiki.creativecommons.org/Considerations_for_licensees
=======================================================================
Creative Commons Attribution-NonCommercial 4.0 International Public
License
By exercising the Licensed Rights (defined below), You accept and agree
to be bound by the terms and conditions of this Creative Commons
Attribution-NonCommercial 4.0 International Public License ("Public
License"). To the extent this Public License may be interpreted as a
contract, You are granted the Licensed Rights in consideration of Your
acceptance of these terms and conditions, and the Licensor grants You
such rights in consideration of benefits the Licensor receives from
making the Licensed Material available under these terms and
conditions.
Section 1 -- Definitions.
a. Adapted Material means material subject to Copyright and Similar
Rights that is derived from or based upon the Licensed Material
and in which the Licensed Material is translated, altered,
arranged, transformed, or otherwise modified in a manner requiring
permission under the Copyright and Similar Rights held by the
Licensor. For purposes of this Public License, where the Licensed
Material is a musical work, performance, or sound recording,
Adapted Material is always produced where the Licensed Material is
synched in timed relation with a moving image.
b. Adapter's License means the license You apply to Your Copyright
and Similar Rights in Your contributions to Adapted Material in
accordance with the terms and conditions of this Public License.
c. Copyright and Similar Rights means copyright and/or similar rights
closely related to copyright including, without limitation,
performance, broadcast, sound recording, and Sui Generis Database
Rights, without regard to how the rights are labeled or
categorized. For purposes of this Public License, the rights
specified in Section 2(b)(1)-(2) are not Copyright and Similar
Rights.
d. Effective Technological Measures means those measures that, in the
absence of proper authority, may not be circumvented under laws
fulfilling obligations under Article 11 of the WIPO Copyright
Treaty adopted on December 20, 1996, and/or similar international
agreements.
e. Exceptions and Limitations means fair use, fair dealing, and/or
any other exception or limitation to Copyright and Similar Rights
that applies to Your use of the Licensed Material.
f. Licensed Material means the artistic or literary work, database,
or other material to which the Licensor applied this Public
License.
g. Licensed Rights means the rights granted to You subject to the
terms and conditions of this Public License, which are limited to
all Copyright and Similar Rights that apply to Your use of the
Licensed Material and that the Licensor has authority to license.
h. Licensor means the individual(s) or entity(ies) granting rights
under this Public License.
i. NonCommercial means not primarily intended for or directed towards
commercial advantage or monetary compensation. For purposes of
this Public License, the exchange of the Licensed Material for
other material subject to Copyright and Similar Rights by digital
file-sharing or similar means is NonCommercial provided there is
no payment of monetary compensation in connection with the
exchange.
j. Share means to provide material to the public by any means or
process that requires permission under the Licensed Rights, such
as reproduction, public display, public performance, distribution,
dissemination, communication, or importation, and to make material
available to the public including in ways that members of the
public may access the material from a place and at a time
individually chosen by them.
k. Sui Generis Database Rights means rights other than copyright
resulting from Directive 96/9/EC of the European Parliament and of
the Council of 11 March 1996 on the legal protection of databases,
as amended and/or succeeded, as well as other essentially
equivalent rights anywhere in the world.
l. You means the individual or entity exercising the Licensed Rights
under this Public License. Your has a corresponding meaning.
Section 2 -- Scope.
a. License grant.
1. Subject to the terms and conditions of this Public License,
the Licensor hereby grants You a worldwide, royalty-free,
non-sublicensable, non-exclusive, irrevocable license to
exercise the Licensed Rights in the Licensed Material to:
a. reproduce and Share the Licensed Material, in whole or
in part, for NonCommercial purposes only; and
b. produce, reproduce, and Share Adapted Material for
NonCommercial purposes only.
2. Exceptions and Limitations. For the avoidance of doubt, where
Exceptions and Limitations apply to Your use, this Public
License does not apply, and You do not need to comply with
its terms and conditions.
3. Term. The term of this Public License is specified in Section
6(a).
4. Media and formats; technical modifications allowed. The
Licensor authorizes You to exercise the Licensed Rights in
all media and formats whether now known or hereafter created,
and to make technical modifications necessary to do so. The
Licensor waives and/or agrees not to assert any right or
authority to forbid You from making technical modifications
necessary to exercise the Licensed Rights, including
technical modifications necessary to circumvent Effective
Technological Measures. For purposes of this Public License,
simply making modifications authorized by this Section 2(a)
(4) never produces Adapted Material.
5. Downstream recipients.
a. Offer from the Licensor -- Licensed Material. Every
recipient of the Licensed Material automatically
receives an offer from the Licensor to exercise the
Licensed Rights under the terms and conditions of this
Public License.
b. No downstream restrictions. You may not offer or impose
any additional or different terms or conditions on, or
apply any Effective Technological Measures to, the
Licensed Material if doing so restricts exercise of the
Licensed Rights by any recipient of the Licensed
Material.
6. No endorsement. Nothing in this Public License constitutes or
may be construed as permission to assert or imply that You
are, or that Your use of the Licensed Material is, connected
with, or sponsored, endorsed, or granted official status by,
the Licensor or others designated to receive attribution as
provided in Section 3(a)(1)(A)(i).
b. Other rights.
1. Moral rights, such as the right of integrity, are not
licensed under this Public License, nor are publicity,
privacy, and/or other similar personality rights; however, to
the extent possible, the Licensor waives and/or agrees not to
assert any such rights held by the Licensor to the limited
extent necessary to allow You to exercise the Licensed
Rights, but not otherwise.
2. Patent and trademark rights are not licensed under this
Public License.
3. To the extent possible, the Licensor waives any right to
collect royalties from You for the exercise of the Licensed
Rights, whether directly or through a collecting society
under any voluntary or waivable statutory or compulsory
licensing scheme. In all other cases the Licensor expressly
reserves any right to collect such royalties, including when
the Licensed Material is used other than for NonCommercial
purposes.
Section 3 -- License Conditions.
Your exercise of the Licensed Rights is expressly made subject to the
following conditions.
a. Attribution.
1. If You Share the Licensed Material (including in modified
form), You must:
a. retain the following if it is supplied by the Licensor
with the Licensed Material:
i. identification of the creator(s) of the Licensed
Material and any others designated to receive
attribution, in any reasonable manner requested by
the Licensor (including by pseudonym if
designated);
ii. a copyright notice;
iii. a notice that refers to this Public License;
iv. a notice that refers to the disclaimer of
warranties;
v. a URI or hyperlink to the Licensed Material to the
extent reasonably practicable;
b. indicate if You modified the Licensed Material and
retain an indication of any previous modifications; and
c. indicate the Licensed Material is licensed under this
Public License, and include the text of, or the URI or
hyperlink to, this Public License.
2. You may satisfy the conditions in Section 3(a)(1) in any
reasonable manner based on the medium, means, and context in
which You Share the Licensed Material. For example, it may be
reasonable to satisfy the conditions by providing a URI or
hyperlink to a resource that includes the required
information.
3. If requested by the Licensor, You must remove any of the
information required by Section 3(a)(1)(A) to the extent
reasonably practicable.
4. If You Share Adapted Material You produce, the Adapter's
License You apply must not prevent recipients of the Adapted
Material from complying with this Public License.
Section 4 -- Sui Generis Database Rights.
Where the Licensed Rights include Sui Generis Database Rights that
apply to Your use of the Licensed Material:
a. for the avoidance of doubt, Section 2(a)(1) grants You the right
to extract, reuse, reproduce, and Share all or a substantial
portion of the contents of the database for NonCommercial purposes
only;
b. if You include all or a substantial portion of the database
contents in a database in which You have Sui Generis Database
Rights, then the database in which You have Sui Generis Database
Rights (but not its individual contents) is Adapted Material; and
c. You must comply with the conditions in Section 3(a) if You Share
all or a substantial portion of the contents of the database.
For the avoidance of doubt, this Section 4 supplements and does not
replace Your obligations under this Public License where the Licensed
Rights include other Copyright and Similar Rights.
Section 5 -- Disclaimer of Warranties and Limitation of Liability.
a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
c. The disclaimer of warranties and limitation of liability provided
above shall be interpreted in a manner that, to the extent
possible, most closely approximates an absolute disclaimer and
waiver of all liability.
Section 6 -- Term and Termination.
a. This Public License applies for the term of the Copyright and
Similar Rights licensed here. However, if You fail to comply with
this Public License, then Your rights under this Public License
terminate automatically.
b. Where Your right to use the Licensed Material has terminated under
Section 6(a), it reinstates:
1. automatically as of the date the violation is cured, provided
it is cured within 30 days of Your discovery of the
violation; or
2. upon express reinstatement by the Licensor.
For the avoidance of doubt, this Section 6(b) does not affect any
right the Licensor may have to seek remedies for Your violations
of this Public License.
c. For the avoidance of doubt, the Licensor may also offer the
Licensed Material under separate terms or conditions or stop
distributing the Licensed Material at any time; however, doing so
will not terminate this Public License.
d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
License.
Section 7 -- Other Terms and Conditions.
a. The Licensor shall not be bound by any additional or different
terms or conditions communicated by You unless expressly agreed.
b. Any arrangements, understandings, or agreements regarding the
Licensed Material not stated herein are separate from and
independent of the terms and conditions of this Public License.
Section 8 -- Interpretation.
a. For the avoidance of doubt, this Public License does not, and
shall not be interpreted to, reduce, limit, restrict, or impose
conditions on any use of the Licensed Material that could lawfully
be made without permission under this Public License.
b. To the extent possible, if any provision of this Public License is
deemed unenforceable, it shall be automatically reformed to the
minimum extent necessary to make it enforceable. If the provision
cannot be reformed, it shall be severed from this Public License
without affecting the enforceability of the remaining terms and
conditions.
c. No term or condition of this Public License will be waived and no
failure to comply consented to unless expressly agreed to by the
Licensor.
d. Nothing in this Public License constitutes or may be interpreted
as a limitation upon, or waiver of, any privileges and immunities
that apply to the Licensor or You, including from the legal
processes of any jurisdiction or authority.
=======================================================================
Creative Commons is not a party to its public
licenses. Notwithstanding, Creative Commons may elect to apply one of
its public licenses to material it publishes and in those instances
will be considered the “Licensor.” The text of the Creative Commons
public licenses is dedicated to the public domain under the CC0 Public
Domain Dedication. Except for the limited purpose of indicating that
material is shared under a Creative Commons public license or as
otherwise permitted by the Creative Commons policies published at
creativecommons.org/policies, Creative Commons does not authorize the
use of the trademark "Creative Commons" or any other trademark or logo
of Creative Commons without its prior written consent including,
without limitation, in connection with any unauthorized modifications
to any of its public licenses or any other arrangements,
understandings, or agreements concerning use of licensed material. For
the avoidance of doubt, this paragraph does not form part of the
public licenses.
Creative Commons may be contacted at creativecommons.org.

172
README.md Normal file
View File

@ -0,0 +1,172 @@
# 🐶 Bark
<a href="http://www.repostatus.org/#active"><img src="http://www.repostatus.org/badges/latest/active.svg" /></a>
[![Twitter](https://img.shields.io/twitter/url/https/twitter.com/OnusFM.svg?style=social&label=Follow%20%40OnusFM)](https://twitter.com/OnusFM)
[![](https://dcbadge.vercel.app/api/server/J2B2vsjKuE?compact=true&style=flat)](https://discord.gg/J2B2vsjKuE)
[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1mnpdHViSiu4VIDGYA3k8aNgXmgOGWgU-?usp=sharing)
[Examples](https://suno-ai.notion.site/Bark-Examples-5edae8b02a604b54a42244ba45ebc2e2) | [Model Card](./model-card.md)
Bark is a transformer-based text-to-audio model created by [Suno](https://suno.ai). It can generate highly realistic multilingual speech, other audio, including music and background noise, and speaker emotions like laughing, sighing and crying. To support the community we give access to pretrained model checkpoints ready for inference.
<p align="center">
<img src="https://user-images.githubusercontent.com/5068315/230698495-cbb1ced9-c911-4c9a-941d-a1a4a1286ac6.png" width="500"></img>
</p>
## 🤖 Usage
[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1mnpdHViSiu4VIDGYA3k8aNgXmgOGWgU-?usp=sharing)
```python
from bark import SAMPLE_RATE, generate_audio
from IPython.display import Audio
text_prompt = """
Hello, my name is Suno. And, uh — and I like pizza. [laughs]
But I also have other interests such as playing tic tac toe.
"""
audio_array = generate_audio(text_prompt)
Audio(audio_array, rate=SAMPLE_RATE)
```
[pizza.webm](https://user-images.githubusercontent.com/5068315/230490503-417e688d-5115-4eee-9550-b46a2b465ee3.webm)
### 🌎 Foreign Language
Bark supports various languages out-of-the-box and automatically determines language from input text. Code-switched text will even realistically use the same voice and add an accent.
```python
text_prompt = """
Buenos días Miguel. Tu colega piensa que tu alemán es extremadamente malo.
But I suppose your english isn't terrible.
"""
audio_array = generate_audio(text_prompt)
```
[miguel.webm](https://user-images.githubusercontent.com/5068315/230684752-10baadfe-1e7c-46a2-8323-43282aef2c8c.webm)
### 🎶 Music
Bark can generate all types of audio, and in principle doesn't see a difference between speech and music. Sometimes it chooses to generate text as music, but you can help it out by adding notes around your lyrics.
```python
text_prompt = """
♪ In the jungle, the mighty jungle, the lion barks tonight ♪
"""
audio_array = generate_audio(text_prompt)
```
[lion.webm](https://user-images.githubusercontent.com/5068315/230684766-97f5ea23-ad99-473c-924b-66b6fab24289.webm)
### 👥 Speaker Prompts
You can provide certain speaker prompts such as NARRATOR, MAN, WOMAN, etc. (Note that these are not always respected, especially if a conflicting audio history prompt is given.)
```python
text_prompt = """
WOMAN: I would like an oatmilk latte please.
MAN: Wow, that's expensive!
"""
audio_array = generate_audio(text_prompt)
```
[latte.webm](https://user-images.githubusercontent.com/5068315/230684864-12d101a1-a726-471d-9d56-d18b108efcb8.webm)
### 🎤 Voice/Audio Cloning
Bark has the capability to fully clone voices as well pick up music, ambience, etc. from input clips. However, to avoid misuse of this technology we limit the audio history prompts to a limited set of Suno-provided, fully synthetic options to choose from.
```python
text_prompt = """
I have a silky smooth voice, and today I will tell you about
the exercise regimen of the common sloth.
"""
audio_array = generate_audio(text_prompt, history_prompt="man-narrator")
```
[sloth.webm](https://user-images.githubusercontent.com/5068315/230684883-a344c619-a560-4ff5-8b99-b4463a34487b.webm)
## 💻 Installation
```
pip install git+https://github.com/suno-ai/bark.git
```
or
```
git clone https://github.com/suno-ai/bark
cd bark && pip install .
```
## 🛠️ Hardware and Inference Speed
Bark has been tested and works on both CPU and GPU (`pytorch 2.0+`, CUDA 11.7 and CUDA 12.0).
Running Bark requires running >100M parameter transformer models.
On modern GPUs and PyTorch nightly, Bark can generate audio in roughly realtime. On older GPUs, default colab, or CPU, inference time might be 10-100x slower.
If you don't have new hardware available or if you want to play with bigger versions of our models, you can also sign up for early access to our Studio [here](https://3os84zs17th.typeform.com/suno-studio).
## ⚙️ Details
Similar to [Vall-E](https://arxiv.org/abs/2301.02111) and some other amazing work in the field, Bark uses GPT-style
models to generate audio from scratch. Different from Vall-E, the initial text prompt is embedded into high-level semantic tokens without the use of phonemes. It can therefore generalize to arbitrary instructions beyond speech that occur in the training data, such as music lyrics, sound effects or other non-speech sounds. A subsequent second model is used to convert the generated semantic tokens into audio codec tokens to generate the full waveform. To enable the community to use Bark via public code we used the fantastic
[EnCodec codec](https://github.com/facebookresearch/encodec) from Facebook to act as an audio representation.
Below is a list of some known non-speech sounds, but we are finding more every day. Please let us know if you find patterns that work particularly well on [Discord](https://discord.gg/DqEx7FGbFP)!
- `[laughter]`
- `[laughs]`
- `[sighs]`
- `[music]`
- `[gasps]`
- `[clears throat]`
- `—` or `...` for hesitations
- `♪` for song lyrics
- capitalization for emphasis of a word
- `MAN/WOMAN:` for bias towards speaker
**Supported Languages**
| Language | Status |
| --- | --- |
| Chinese (Mandarin) | ✅ |
| English | ✅ |
| French | ✅ |
| German | ✅ |
| Hindi | ✅ |
| Italian | ✅ |
| Japanese | ✅ |
| Korean | ✅ |
| Polish | ✅ |
| Portuguese | ✅ |
| Russian | ✅ |
| Spanish | ✅ |
| Turkish | ✅ |
| Arabic | Coming soon! |
| Bengali | Coming soon! |
| Telugu | Coming soon! |
## 🙏 Appreciation
- [nanoGPT](https://github.com/karpathy/nanoGPT) for a dead-simple and blazing fast implementation of gpt-style models
- [EnCodec](https://github.com/facebookresearch/encodec) for a state-of-the-art implementation of a fantastic audio codec
- [AudioLM](https://github.com/lucidrains/audiolm-pytorch) for very related training and inference code
- [Vall-E](https://arxiv.org/abs/2301.02111), [AudioLM](https://arxiv.org/abs/2209.03143) and many other ground-breaking papers that enabled the development of Bark
## © License
Bark is licensed under a non-commercial CC-BY 4.0 NC. The Suno models themselves may be used commercially. However, this version of Bark uses `EnCodec` as a neural codec backend, which is licensed under a [non-commercial license](https://github.com/facebookresearch/encodec/blob/main/LICENSE).
Please contact us at `bark@suno.ai` if you need access to a larger version of the model and/or a version of the model you can use commercially.
## 📱 Community
- [Twitter](https://twitter.com/OnusFM)
- [Discord](https://discord.gg/J2B2vsjKuE)
## 🎧 Suno Studio (Early Access)
We’re developing a web interface for our models, including Bark.
You can sign up for early access [here](https://3os84zs17th.typeform.com/suno-studio).

2
bark/__init__.py Normal file
View File

@ -0,0 +1,2 @@
from .api import generate_audio, text_to_semantic, semantic_to_waveform
from .generation import SAMPLE_RATE, preload_models

BIN
bark/__pycache__/__init__.cpython-38.pyc Normal file
View File

Binary file not shown.

BIN
bark/__pycache__/api.cpython-38.pyc Normal file
View File

Binary file not shown.

BIN
bark/__pycache__/generation.cpython-38.pyc Normal file
View File

Binary file not shown.

BIN
bark/__pycache__/model.cpython-38.pyc Normal file
View File

Binary file not shown.

BIN
bark/__pycache__/model_fine.cpython-38.pyc Normal file
View File

Binary file not shown.

79
bark/api.py Normal file
View File

@ -0,0 +1,79 @@
from typing import Optional
import numpy as np
from .generation import codec_decode, generate_coarse, generate_fine, generate_text_semantic
def text_to_semantic(
text: str,
history_prompt: Optional[str] = None,
temp: float = 0.7,
):
"""Generate semantic array from text.
Args:
text: text to be turned into audio
history_prompt: history choice for audio cloning
temp: generation temperature (1.0 more diverse, 0.0 more conservative)
Returns:
numpy semantic array to be fed into `semantic_to_waveform`
"""
x_semantic = generate_text_semantic(
text,
history_prompt=history_prompt,
temp=temp,
)
return x_semantic
def semantic_to_waveform(
semantic_tokens: np.ndarray,
history_prompt: Optional[str] = None,
temp: float = 0.7,
):
"""Generate audio array from semantic input.
Args:
semantic_tokens: semantic token output from `text_to_semantic`
history_prompt: history choice for audio cloning
temp: generation temperature (1.0 more diverse, 0.0 more conservative)
Returns:
numpy audio array at sample frequency 24khz
"""
x_coarse_gen = generate_coarse(
semantic_tokens,
history_prompt=history_prompt,
temp=temp,
)
x_fine_gen = generate_fine(
x_coarse_gen,
history_prompt=history_prompt,
temp=0.5,
)
audio_arr = codec_decode(x_fine_gen)
return audio_arr
def generate_audio(
text: str,
history_prompt: Optional[str] = None,
text_temp: float = 0.7,
waveform_temp: float = 0.7,
):
"""Generate audio array from input text.
Args:
text: text to be turned into audio
history_prompt: history choice for audio cloning
text_temp: generation temperature (1.0 more diverse, 0.0 more conservative)
waveform_temp: generation temperature (1.0 more diverse, 0.0 more conservative)
Returns:
numpy audio array at sample frequency 24khz
"""
x_semantic = text_to_semantic(text, history_prompt=history_prompt, temp=text_temp)
audio_arr = semantic_to_waveform(x_semantic, history_prompt=history_prompt, temp=waveform_temp)
return audio_arr

BIN
bark/assets/prompts/brylcream.npz Normal file
View File

Binary file not shown.

BIN
bark/assets/prompts/es-woman.npz Normal file
View File

Binary file not shown.

BIN
bark/assets/prompts/man-narrator.npz Normal file
View File

Binary file not shown.

693
bark/generation.py Normal file
View File

@ -0,0 +1,693 @@
import contextlib
import hashlib
import os
import re
import requests
import sys
from encodec import EncodecModel
import funcy
import logging
import numpy as np
from scipy.special import softmax
import torch
import torch.nn.functional as F
import tqdm
from transformers import BertTokenizer
from .model import GPTConfig, GPT
from .model_fine import FineGPT, FineGPTConfig
if (
torch.cuda.is_available() and
hasattr(torch.cuda, "amp") and
hasattr(torch.cuda.amp, "autocast") and
torch.cuda.is_bf16_supported()
):
autocast = funcy.partial(torch.cuda.amp.autocast, dtype=torch.bfloat16)
else:
@contextlib.contextmanager
def autocast():
yield
# hold models in global scope to lazy load
global models
models = {}
CONTEXT_WINDOW_SIZE = 1024
SEMANTIC_RATE_HZ = 49.9
SEMANTIC_VOCAB_SIZE = 10_000
CODEBOOK_SIZE = 1024
N_COARSE_CODEBOOKS = 2
N_FINE_CODEBOOKS = 8
COARSE_RATE_HZ = 75
SAMPLE_RATE = 24_000
ALLOWED_PROMPTS = (
"brylcream",
"es-woman",
"man-narrator",
)
logger = logging.getLogger(__name__)
CUR_PATH = os.path.dirname(os.path.abspath(__file__))
default_cache_dir = os.path.join(os.path.expanduser("~"), ".cache")
CACHE_DIR = os.path.join(os.getenv("XDG_CACHE_HOME", default_cache_dir), "suno", "bark_v0")
os.makedirs(CACHE_DIR, exist_ok=True)
REMOTE_BASE_URL = "http://s3.amazonaws.com/suno-public/bark/models/v0/"
REMOTE_MODEL_PATHS = {
"text": os.environ.get("SUNO_TEXT_MODEL_PATH", os.path.join(REMOTE_BASE_URL, "text.pt")),
"coarse": os.environ.get("SUNO_COARSE_MODEL_PATH", os.path.join(REMOTE_BASE_URL, "coarse.pt")),
"fine": os.environ.get("SUNO_FINE_MODEL_PATH", os.path.join(REMOTE_BASE_URL, "fine.pt")),
}
def _compute_md5(s):
m = hashlib.md5()
m.update(s.encode("utf-8"))
return m.hexdigest()
def _get_ckpt_path(model_type):
model_name = _compute_md5(REMOTE_MODEL_PATHS[model_type])
return os.path.join(CACHE_DIR, f"{model_name}.pt")
S3_BUCKET_PATH_RE = r"s3\:\/\/(.+?)\/"
def _parse_s3_filepath(s3_filepath):
bucket_name = re.search(S3_BUCKET_PATH_RE, s3_filepath).group(1)
rel_s3_filepath = re.sub(S3_BUCKET_PATH_RE, "", s3_filepath)
return bucket_name, rel_s3_filepath
def _download(from_s3_path, to_local_path):
response = requests.get(from_s3_path, stream=True)
total_size_in_bytes = int(response.headers.get('content-length', 0))
block_size = 1024 # 1 Kibibyte
progress_bar = tqdm.tqdm(total=total_size_in_bytes, unit='iB', unit_scale=True)
with open(to_local_path, "wb") as file:
for data in response.iter_content(block_size):
progress_bar.update(len(data))
file.write(data)
progress_bar.close()
if total_size_in_bytes != 0 and progress_bar.n != total_size_in_bytes:
raise ValueError("ERROR, something went wrong")
class InferenceContext:
def __init__(self, benchmark=False):
# we can't expect inputs to be the same length, so disable benchmarking by default
self._chosen_cudnn_benchmark = benchmark
self._cudnn_benchmark = None
def __enter__(self):
self._cudnn_benchmark = torch.backends.cudnn.benchmark
torch.backends.cudnn.benchmark = self._chosen_cudnn_benchmark
def __exit__(self, exc_type, exc_value, exc_traceback):
torch.backends.cudnn.benchmark = self._cudnn_benchmark
if torch.cuda.is_available():
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
@contextlib.contextmanager
def _inference_mode():
with InferenceContext(), torch.inference_mode(), torch.no_grad(), autocast():
yield
def _clear_cuda_cache():
if torch.cuda.is_available():
torch.cuda.empty_cache()
torch.cuda.synchronize()
def clean_models(model_key=None):
global models
model_keys = [model_key] if model_key is not None else models.keys()
for k in model_keys:
if k in models:
del models[k]
_clear_cuda_cache()
def _load_model(ckpt_path, device, model_type="text"):
if "cuda" not in device:
logger.warning("No GPU being used. Careful, Inference might be extremely slow!")
if model_type == "text":
ConfigClass = GPTConfig
ModelClass = GPT
elif model_type == "coarse":
ConfigClass = GPTConfig
ModelClass = GPT
elif model_type == "fine":
ConfigClass = FineGPTConfig
ModelClass = FineGPT
else:
raise NotImplementedError()
if not os.path.exists(ckpt_path):
print(f"{model_type} model not found, downloading...")
_download(REMOTE_MODEL_PATHS[model_type], ckpt_path)
checkpoint = torch.load(ckpt_path, map_location=device)
# this is a hack
model_args = checkpoint["model_args"]
if "input_vocab_size" not in model_args:
model_args["input_vocab_size"] = model_args["vocab_size"]
model_args["output_vocab_size"] = model_args["vocab_size"]
del model_args["vocab_size"]
gptconf = ConfigClass(**checkpoint["model_args"])
model = ModelClass(gptconf)
state_dict = checkpoint["model"]
# fixup checkpoint
unwanted_prefix = "_orig_mod."
for k, v in list(state_dict.items()):
if k.startswith(unwanted_prefix):
state_dict[k[len(unwanted_prefix) :]] = state_dict.pop(k)
extra_keys = set(state_dict.keys()) - set(model.state_dict().keys())
extra_keys = set([k for k in extra_keys if not k.endswith(".attn.bias")])
missing_keys = set(model.state_dict().keys()) - set(state_dict.keys())
missing_keys = set([k for k in missing_keys if not k.endswith(".attn.bias")])
if len(extra_keys) != 0:
raise ValueError(f"extra keys found: {extra_keys}")
if len(missing_keys) != 0:
raise ValueError(f"missing keys: {missing_keys}")
model.load_state_dict(state_dict, strict=False)
n_params = model.get_num_params()
val_loss = checkpoint["best_val_loss"].item()
print(f"model loaded: {round(n_params/1e6,1)}M params, {round(val_loss,3)} loss")
model.eval()
model.to(device)
del checkpoint, state_dict
_clear_cuda_cache()
if model_type == "text":
tokenizer = BertTokenizer.from_pretrained("bert-base-multilingual-cased")
return {
"model": model,
"tokenizer": tokenizer,
}
return model
def _load_codec_model(device):
model = EncodecModel.encodec_model_24khz()
model.set_target_bandwidth(6.0)
model.eval()
model.to(device)
_clear_cuda_cache()
return model
def load_model(ckpt_path=None, use_gpu=True, force_reload=False, model_type="text"):
_load_model_f = funcy.partial(_load_model, model_type=model_type)
if model_type not in ("text", "coarse", "fine"):
raise NotImplementedError()
global models
if torch.cuda.device_count() == 0 or not use_gpu:
device = "cpu"
else:
device = "cuda"
model_key = str(device) + f"__{model_type}"
if model_key not in models or force_reload:
if ckpt_path is None:
ckpt_path = _get_ckpt_path(model_type)
clean_models(model_key=model_key)
model = _load_model_f(ckpt_path, device)
models[model_key] = model
return models[model_key]
def load_codec_model(use_gpu=True, force_reload=False):
global models
if torch.cuda.device_count() == 0 or not use_gpu:
device = "cpu"
else:
device = "cuda"
model_key = str(device) + f"__codec"
if model_key not in models or force_reload:
clean_models(model_key=model_key)
model = _load_codec_model(device)
models[model_key] = model
return models[model_key]
def preload_models(text_ckpt_path=None, coarse_ckpt_path=None, fine_ckpt_path=None, use_gpu=True):
_ = load_model(
ckpt_path=text_ckpt_path, model_type="text", use_gpu=use_gpu, force_reload=True
)
_ = load_model(
ckpt_path=coarse_ckpt_path, model_type="coarse", use_gpu=use_gpu, force_reload=True
)
_ = load_model(
ckpt_path=fine_ckpt_path, model_type="fine", use_gpu=use_gpu, force_reload=True
)
_ = load_codec_model(use_gpu=use_gpu, force_reload=True)
####
# Generation Functionality
####
def _tokenize(tokenizer, text):
return tokenizer.encode(text, add_special_tokens=False)
def _detokenize(tokenizer, enc_text):
return tokenizer.decode(enc_text)
def _normalize_whitespace(text):
return re.sub(r"\s+", " ", text).strip()
TEXT_ENCODING_OFFSET = 10_048
SEMANTIC_PAD_TOKEN = 10_000
TEXT_PAD_TOKEN = 129_595
SEMANTIC_INFER_TOKEN = 129_599
def generate_text_semantic(
text,
history_prompt=None,
temp=0.7,
top_k=None,
top_p=None,
use_gpu=True,
silent=False,
min_eos_p=0.2,
max_gen_duration_s=None,
allow_early_stop=True,
model=None,
):
"""Generate semantic tokens from text."""
assert isinstance(text, str)
text = _normalize_whitespace(text)
assert len(text.strip()) > 0
if history_prompt is not None:
assert (history_prompt in ALLOWED_PROMPTS)
semantic_history = np.load(
os.path.join(CUR_PATH, "assets", "prompts", f"{history_prompt}.npz")
)["text"]
assert (
isinstance(semantic_history, np.ndarray)
and len(semantic_history.shape) == 1
and len(semantic_history) > 0
and semantic_history.min() >= 0
and semantic_history.max() <= SEMANTIC_VOCAB_SIZE - 1
)
else:
semantic_history = None
model_container = load_model(use_gpu=use_gpu, model_type="text")
if model is None:
model = model_container["model"]
tokenizer = model_container["tokenizer"]
encoded_text = np.array(_tokenize(tokenizer, text)) + TEXT_ENCODING_OFFSET
device = "cuda" if use_gpu and torch.cuda.device_count() > 0 else "cpu"
if len(encoded_text) > 256:
p = round((len(encoded_text) - 256) / len(encoded_text) * 100, 1)
print(f"warning, text too long, lopping of last {p}%")
encoded_text = encoded_text[:256]
encoded_text = np.pad(
encoded_text,
(0, 256 - len(encoded_text)),
constant_values=TEXT_PAD_TOKEN,
mode="constant",
)
if semantic_history is not None:
semantic_history = semantic_history.astype(np.int64)
# lop off if history is too long, pad if needed
semantic_history = semantic_history[-256:]
semantic_history = np.pad(
semantic_history,
(0, 256 - len(semantic_history)),
constant_values=SEMANTIC_PAD_TOKEN,
mode="constant",
)
else:
semantic_history = np.array([SEMANTIC_PAD_TOKEN] * 256)
x = torch.from_numpy(
np.hstack([encoded_text, semantic_history, np.array([SEMANTIC_INFER_TOKEN])]).astype(np.int64)
)[None]
assert x.shape[1] == 256 + 256 + 1
with _inference_mode():
x = x.to(device)
n_tot_steps = 768
# custom tqdm updates since we don't know when eos will occur
pbar = tqdm.tqdm(disable=silent, total=100)
pbar_state = 0
tot_generated_duration_s = 0
for n in range(n_tot_steps):
logits = model(x, merge_context=True)
relevant_logits = logits[0, 0, :SEMANTIC_VOCAB_SIZE]
if allow_early_stop:
relevant_logits = torch.hstack(
(relevant_logits, logits[0, 0, [SEMANTIC_PAD_TOKEN]]) # eos
)
if top_p is not None:
# faster to convert to numpy
logits_device = relevant_logits.device
logits_dtype = relevant_logits.type()
relevant_logits = relevant_logits.detach().cpu().type(torch.float32).numpy()
sorted_indices = np.argsort(relevant_logits)[::-1]
sorted_logits = relevant_logits[sorted_indices]
cumulative_probs = np.cumsum(softmax(sorted_logits))
sorted_indices_to_remove = cumulative_probs > top_p
sorted_indices_to_remove[1:] = sorted_indices_to_remove[:-1].copy()
sorted_indices_to_remove[0] = False
relevant_logits[sorted_indices[sorted_indices_to_remove]] = -np.inf
relevant_logits = torch.from_numpy(relevant_logits)
relevant_logits = relevant_logits.to(logits_device).type(logits_dtype)
if top_k is not None:
v, _ = torch.topk(relevant_logits, min(top_k, relevant_logits.size(-1)))
relevant_logits[relevant_logits < v[-1]] = -float("Inf")
probs = F.softmax(relevant_logits / temp, dim=-1)
item_next = torch.multinomial(probs, num_samples=1)
if allow_early_stop and (
item_next == SEMANTIC_VOCAB_SIZE
or (min_eos_p is not None and probs[-1] >= min_eos_p)
):
# eos found, so break
pbar.update(100 - pbar_state)
break
x = torch.cat((x, item_next[None]), dim=1)
tot_generated_duration_s += 1 / SEMANTIC_RATE_HZ
if max_gen_duration_s is not None and tot_generated_duration_s > max_gen_duration_s:
pbar.update(100 - pbar_state)
break
if n == n_tot_steps - 1:
pbar.update(100 - pbar_state)
break
del logits, relevant_logits, probs, item_next
req_pbar_state = np.min([100, int(round(100 * n / n_tot_steps))])
if req_pbar_state > pbar_state:
pbar.update(req_pbar_state - pbar_state)
pbar_state = req_pbar_state
pbar.close()
out = x.detach().cpu().numpy().squeeze()[256 + 256 + 1 :]
assert all(0 <= out) and all(out < SEMANTIC_VOCAB_SIZE)
_clear_cuda_cache()
return out
def _flatten_codebooks(arr, offset_size=CODEBOOK_SIZE):
assert len(arr.shape) == 2
arr = arr.copy()
if offset_size is not None:
for n in range(1, arr.shape[0]):
arr[n, :] += offset_size * n
flat_arr = arr.ravel("F")
return flat_arr
COARSE_SEMANTIC_PAD_TOKEN = 12_048
COARSE_INFER_TOKEN = 12_050
def generate_coarse(
x_semantic,
history_prompt=None,
temp=0.7,
top_k=None,
top_p=None,
use_gpu=True,
silent=False,
max_coarse_history=630, # min 60 (faster), max 630 (more context)
sliding_window_len=60,
model=None,
):
"""Generate coarse audio codes from semantic tokens."""
assert (
isinstance(x_semantic, np.ndarray)
and len(x_semantic.shape) == 1
and len(x_semantic) > 0
and x_semantic.min() >= 0
and x_semantic.max() <= SEMANTIC_VOCAB_SIZE - 1
)
assert 60 <= max_coarse_history <= 630
assert max_coarse_history + sliding_window_len <= 1024 - 256
semantic_to_coarse_ratio = COARSE_RATE_HZ / SEMANTIC_RATE_HZ * N_COARSE_CODEBOOKS
max_semantic_history = int(np.floor(max_coarse_history / semantic_to_coarse_ratio))
if history_prompt is not None:
assert (history_prompt in ALLOWED_PROMPTS)
x_history = np.load(
os.path.join(CUR_PATH, "assets", "prompts", f"{history_prompt}.npz")
)
x_semantic_history = x_history["coarse_1"]
x_coarse_history = x_history["coarse_2"]
assert (
isinstance(x_semantic_history, np.ndarray)
and len(x_semantic_history.shape) == 1
and len(x_semantic_history) > 0
and x_semantic_history.min() >= 0
and x_semantic_history.max() <= SEMANTIC_VOCAB_SIZE - 1
and isinstance(x_coarse_history, np.ndarray)
and len(x_coarse_history.shape) == 2
and x_coarse_history.shape[0] == N_COARSE_CODEBOOKS
and x_coarse_history.shape[-1] >= 0
and x_coarse_history.min() >= 0
and x_coarse_history.max() <= CODEBOOK_SIZE - 1
and (
round(x_coarse_history.shape[-1] / len(x_semantic_history), 1)
== round(semantic_to_coarse_ratio / N_COARSE_CODEBOOKS, 1)
)
)
x_coarse_history = _flatten_codebooks(x_coarse_history) + SEMANTIC_VOCAB_SIZE
# trim histories correctly
n_semantic_hist_provided = np.min(
[
max_semantic_history,
len(x_semantic_history) - len(x_semantic_history) % 2,
int(np.floor(len(x_coarse_history) / semantic_to_coarse_ratio)),
]
)
n_coarse_hist_provided = int(round(n_semantic_hist_provided * semantic_to_coarse_ratio))
x_semantic_history = x_semantic_history[-n_semantic_hist_provided:].astype(np.int32)
x_coarse_history = x_coarse_history[-n_coarse_hist_provided:].astype(np.int32)
# TODO: bit of a hack for time alignment (sounds better)
x_coarse_history = x_coarse_history[:-2]
else:
x_semantic_history = np.array([], dtype=np.int32)
x_coarse_history = np.array([], dtype=np.int32)
if model is None:
model = load_model(use_gpu=use_gpu, model_type="coarse")
device = "cuda" if use_gpu and torch.cuda.device_count() > 0 else "cpu"
# start loop
n_steps = int(
round(
np.floor(len(x_semantic) * semantic_to_coarse_ratio / N_COARSE_CODEBOOKS)
* N_COARSE_CODEBOOKS
)
)
assert n_steps > 0 and n_steps % N_COARSE_CODEBOOKS == 0
x_semantic = np.hstack([x_semantic_history, x_semantic]).astype(np.int32)
x_coarse = x_coarse_history.astype(np.int32)
base_semantic_idx = len(x_semantic_history)
with _inference_mode():
x_semantic_in = torch.from_numpy(x_semantic)[None].to(device)
x_coarse_in = torch.from_numpy(x_coarse)[None].to(device)
n_window_steps = int(np.ceil(n_steps / sliding_window_len))
n_step = 0
for _ in tqdm.tqdm(range(n_window_steps), total=n_window_steps, disable=silent):
semantic_idx = base_semantic_idx + int(round(n_step / semantic_to_coarse_ratio))
# pad from right side
x_in = x_semantic_in[:, np.max([0, semantic_idx - max_semantic_history]) :]
x_in = x_in[:, :256]
x_in = F.pad(
x_in,
(0, 256 - x_in.shape[-1]),
"constant",
COARSE_SEMANTIC_PAD_TOKEN,
)
x_in = torch.hstack(
[
x_in,
torch.tensor([COARSE_INFER_TOKEN])[None].to(device),
x_coarse_in[:, -max_coarse_history:],
]
)
for _ in range(sliding_window_len):
if n_step >= n_steps:
continue
is_major_step = n_step % N_COARSE_CODEBOOKS == 0
logits = model(x_in)
logit_start_idx = (
SEMANTIC_VOCAB_SIZE + (1 - int(is_major_step)) * CODEBOOK_SIZE
)
logit_end_idx = (
SEMANTIC_VOCAB_SIZE + (2 - int(is_major_step)) * CODEBOOK_SIZE
)
relevant_logits = logits[0, 0, logit_start_idx:logit_end_idx]
if top_p is not None:
# faster to convert to numpy
logits_device = relevant_logits.device
logits_dtype = relevant_logits.type()
relevant_logits = relevant_logits.detach().cpu().type(torch.float32).numpy()
sorted_indices = np.argsort(relevant_logits)[::-1]
sorted_logits = relevant_logits[sorted_indices]
cumulative_probs = np.cumsum(softmax(sorted_logits))
sorted_indices_to_remove = cumulative_probs > top_p
sorted_indices_to_remove[1:] = sorted_indices_to_remove[:-1].copy()
sorted_indices_to_remove[0] = False
relevant_logits[sorted_indices[sorted_indices_to_remove]] = -np.inf
relevant_logits = torch.from_numpy(relevant_logits)
relevant_logits = relevant_logits.to(logits_device).type(logits_dtype)
if top_k is not None:
v, _ = torch.topk(relevant_logits, min(top_k, relevant_logits.size(-1)))
relevant_logits[relevant_logits < v[-1]] = -float("Inf")
probs = F.softmax(relevant_logits / temp, dim=-1)
item_next = torch.multinomial(probs, num_samples=1)
item_next += logit_start_idx
x_coarse_in = torch.cat((x_coarse_in, item_next[None]), dim=1)
x_in = torch.cat((x_in, item_next[None]), dim=1)
del logits, relevant_logits, probs, item_next
n_step += 1
del x_in
del x_semantic_in
gen_coarse_arr = x_coarse_in.detach().cpu().numpy().squeeze()[len(x_coarse_history) :]
del x_coarse_in
assert len(gen_coarse_arr) == n_steps
gen_coarse_audio_arr = gen_coarse_arr.reshape(-1, N_COARSE_CODEBOOKS).T - SEMANTIC_VOCAB_SIZE
for n in range(1, N_COARSE_CODEBOOKS):
gen_coarse_audio_arr[n, :] -= n * CODEBOOK_SIZE
_clear_cuda_cache()
return gen_coarse_audio_arr
def generate_fine(
x_coarse_gen,
history_prompt=None,
temp=0.5,
use_gpu=True,
silent=True,
model=None,
):
"""Generate full audio codes from coarse audio codes."""
assert (
isinstance(x_coarse_gen, np.ndarray)
and len(x_coarse_gen.shape) == 2
and 1 <= x_coarse_gen.shape[0] <= N_FINE_CODEBOOKS - 1
and x_coarse_gen.shape[1] > 0
and x_coarse_gen.min() >= 0
and x_coarse_gen.max() <= CODEBOOK_SIZE - 1
)
if history_prompt is not None:
assert (history_prompt in ALLOWED_PROMPTS)
x_fine_history = np.load(
os.path.join(CUR_PATH, "assets", "prompts", f"{history_prompt}.npz")
)["fine"]
assert (
isinstance(x_fine_history, np.ndarray)
and len(x_fine_history.shape) == 2
and x_fine_history.shape[0] == N_FINE_CODEBOOKS
and x_fine_history.shape[1] >= 0
and x_fine_history.min() >= 0
and x_fine_history.max() <= CODEBOOK_SIZE - 1
)
else:
x_fine_history = None
n_coarse = x_coarse_gen.shape[0]
if model is None:
model = load_model(use_gpu=use_gpu, model_type="fine")
device = "cuda" if use_gpu and torch.cuda.device_count() > 0 else "cpu"
# make input arr
in_arr = np.vstack(
[
x_coarse_gen,
np.zeros((N_FINE_CODEBOOKS - n_coarse, x_coarse_gen.shape[1]))
+ CODEBOOK_SIZE, # padding
]
).astype(np.int32)
# prepend history if available (max 512)
if x_fine_history is not None:
x_fine_history = x_fine_history.astype(np.int32)
in_arr = np.hstack(
[
x_fine_history[:, -512:].astype(np.int32),
in_arr,
]
)
n_history = x_fine_history[:, -512:].shape[1]
else:
n_history = 0
n_remove_from_end = 0
# need to pad if too short (since non-causal model)
if in_arr.shape[1] < 1024:
n_remove_from_end = 1024 - in_arr.shape[1]
in_arr = np.hstack(
[
in_arr,
np.zeros((N_FINE_CODEBOOKS, n_remove_from_end), dtype=np.int32) + CODEBOOK_SIZE,
]
)
# we can be lazy about fractional loop and just keep overwriting codebooks
n_loops = np.max([0, int(np.ceil((x_coarse_gen.shape[1] - (1024 - n_history)) / 512))]) + 1
with _inference_mode():
in_arr = torch.tensor(in_arr.T).to(device)
for n in tqdm.tqdm(range(n_loops), disable=silent):
start_idx = np.min([n * 512, in_arr.shape[0] - 1024])
start_fill_idx = np.min([n_history + n * 512, in_arr.shape[0] - 512])
rel_start_fill_idx = start_fill_idx - start_idx
in_buffer = in_arr[start_idx : start_idx + 1024, :][None]
for nn in range(n_coarse, N_FINE_CODEBOOKS):
logits = model(nn, in_buffer)
if temp is None:
relevant_logits = logits[0, rel_start_fill_idx:, :CODEBOOK_SIZE]
codebook_preds = torch.argmax(relevant_logits, -1)
else:
relevant_logits = logits[0, :, :CODEBOOK_SIZE] / temp
probs = F.softmax(relevant_logits, dim=-1)
codebook_preds = torch.hstack(
[
torch.multinomial(probs[n], num_samples=1)
for n in range(rel_start_fill_idx, 1024)
]
)
in_buffer[0, rel_start_fill_idx:, nn] = codebook_preds
del logits, codebook_preds
# transfer over info into model_in and convert to numpy
for nn in range(n_coarse, N_FINE_CODEBOOKS):
in_arr[
start_fill_idx : start_fill_idx + (1024 - rel_start_fill_idx), nn
] = in_buffer[0, rel_start_fill_idx:, nn]
del in_buffer
gen_fine_arr = in_arr.detach().cpu().numpy().squeeze().T
del in_arr
gen_fine_arr = gen_fine_arr[:, n_history:]
if n_remove_from_end > 0:
gen_fine_arr = gen_fine_arr[:, :-n_remove_from_end]
assert gen_fine_arr.shape[-1] == x_coarse_gen.shape[-1]
_clear_cuda_cache()
return gen_fine_arr
def codec_decode(fine_tokens, model=None, use_gpu=True):
"""Turn quantized audio codes into audio array using encodec."""
if model is None:
model = load_codec_model(use_gpu=use_gpu)
device = "cuda" if use_gpu and torch.cuda.device_count() > 0 else "cpu"
arr = torch.from_numpy(fine_tokens)[None]
arr = arr.to(device)
arr = arr.transpose(0, 1)
emb = model.quantizer.decode(arr)
out = model.decoder(emb)
audio_arr = out.detach().cpu().numpy().squeeze()
del arr, emb, out
return audio_arr

174
bark/model.py Normal file
View File

@ -0,0 +1,174 @@
"""
Much of this code is adapted from Andrej Karpathy's NanoGPT
(https://github.com/karpathy/nanoGPT)
"""
import math
from dataclasses import dataclass
import torch
import torch.nn as nn
from torch.nn import functional as F
class LayerNorm(nn.Module):
""" LayerNorm but with an optional bias. PyTorch doesn't support simply bias=False """
def __init__(self, ndim, bias):
super().__init__()
self.weight = nn.Parameter(torch.ones(ndim))
self.bias = nn.Parameter(torch.zeros(ndim)) if bias else None
def forward(self, input):
return F.layer_norm(input, self.weight.shape, self.weight, self.bias, 1e-5)
class CausalSelfAttention(nn.Module):
def __init__(self, config):
super().__init__()
assert config.n_embd % config.n_head == 0
# key, query, value projections for all heads, but in a batch
self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias)
# output projection
self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)
# regularization
self.attn_dropout = nn.Dropout(config.dropout)
self.resid_dropout = nn.Dropout(config.dropout)
self.n_head = config.n_head
self.n_embd = config.n_embd
self.dropout = config.dropout
# flash attention make GPU go brrrrr but support is only in PyTorch nightly and still a bit scary
self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention')
if not self.flash:
# print("WARNING: using slow attention. Flash Attention atm needs PyTorch nightly and dropout=0.0")
# causal mask to ensure that attention is only applied to the left in the input sequence
self.register_buffer("bias", torch.tril(torch.ones(config.block_size, config.block_size))
.view(1, 1, config.block_size, config.block_size))
def forward(self, x):
B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)
# calculate query, key, values for all heads in batch and move head forward to be the batch dim
q, k ,v = self.c_attn(x).split(self.n_embd, dim=2)
k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
# causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)
if self.flash:
# efficient attention using Flash Attention CUDA kernels
y = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=self.dropout, is_causal=True)
else:
# manual implementation of attention
att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
att = att.masked_fill(self.bias[:,:,:T,:T] == 0, float('-inf'))
att = F.softmax(att, dim=-1)
att = self.attn_dropout(att)
y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side
# output projection
y = self.resid_dropout(self.c_proj(y))
return y
class MLP(nn.Module):
def __init__(self, config):
super().__init__()
self.c_fc = nn.Linear(config.n_embd, 4 * config.n_embd, bias=config.bias)
self.c_proj = nn.Linear(4 * config.n_embd, config.n_embd, bias=config.bias)
self.dropout = nn.Dropout(config.dropout)
self.gelu = nn.GELU()
def forward(self, x):
x = self.c_fc(x)
x = self.gelu(x)
x = self.c_proj(x)
x = self.dropout(x)
return x
class Block(nn.Module):
def __init__(self, config, layer_idx):
super().__init__()
self.ln_1 = LayerNorm(config.n_embd, bias=config.bias)
self.attn = CausalSelfAttention(config)
self.ln_2 = LayerNorm(config.n_embd, bias=config.bias)
self.mlp = MLP(config)
self.layer_idx = layer_idx
def forward(self, x):
x = x + self.attn(self.ln_1(x))
x = x + self.mlp(self.ln_2(x))
return x
@dataclass
class GPTConfig:
block_size: int = 1024
input_vocab_size: int = 10_048
output_vocab_size: int = 10_048
n_layer: int = 12
n_head: int = 12
n_embd: int = 768
dropout: float = 0.0
bias: bool = True # True: bias in Linears and LayerNorms, like GPT-2. False: a bit better and faster
class GPT(nn.Module):
def __init__(self, config):
super().__init__()
assert config.input_vocab_size is not None
assert config.output_vocab_size is not None
assert config.block_size is not None
self.config = config
self.transformer = nn.ModuleDict(dict(
wte = nn.Embedding(config.input_vocab_size, config.n_embd),
wpe = nn.Embedding(config.block_size, config.n_embd),
drop = nn.Dropout(config.dropout),
h = nn.ModuleList([Block(config, idx) for idx in range(config.n_layer)]),
ln_f = LayerNorm(config.n_embd, bias=config.bias),
))
self.lm_head = nn.Linear(config.n_embd, config.output_vocab_size, bias=False)
def get_num_params(self, non_embedding=True):
"""
Return the number of parameters in the model.
For non-embedding count (default), the position embeddings get subtracted.
The token embeddings would too, except due to the parameter sharing these
params are actually used as weights in the final layer, so we include them.
"""
n_params = sum(p.numel() for p in self.parameters())
if non_embedding:
n_params -= self.transformer.wte.weight.numel()
n_params -= self.transformer.wpe.weight.numel()
return n_params
def forward(self, idx, merge_context=False):
device = idx.device
b, t = idx.size()
if merge_context:
assert(idx.shape[1] >= 256+256+1)
t = idx.shape[1] - 256
else:
assert t <= self.config.block_size, f"Cannot forward sequence of length {t}, block size is only {self.config.block_size}"
# forward the GPT model itself
if merge_context:
tok_emb = torch.cat([
self.transformer.wte(idx[:,:256]) + self.transformer.wte(idx[:,256:256+256]),
self.transformer.wte(idx[:,256+256:])
], dim=1)
else:
tok_emb = self.transformer.wte(idx) # token embeddings of shape (b, t, n_embd)
pos = torch.arange(0, t, dtype=torch.long, device=device).unsqueeze(0) # shape (1, t)
pos_emb = self.transformer.wpe(pos) # position embeddings of shape (1, t, n_embd)
x = self.transformer.drop(tok_emb + pos_emb)
for block in self.transformer.h:
x = block(x)
x = self.transformer.ln_f(x)
# inference-time mini-optimization: only forward the lm_head on the very last position
logits = self.lm_head(x[:, [-1], :]) # note: using list [-1] to preserve the time dim
return logits

149
bark/model_fine.py Normal file
View File

@ -0,0 +1,149 @@
"""
Much of this code is adapted from Andrej Karpathy's NanoGPT
(https://github.com/karpathy/nanoGPT)
"""
from dataclasses import dataclass
import math
import torch
import torch.nn as nn
from torch.nn import functional as F
from .model import GPT, GPTConfig, MLP
class NonCausalSelfAttention(nn.Module):
def __init__(self, config):
super().__init__()
assert config.n_embd % config.n_head == 0
# key, query, value projections for all heads, but in a batch
self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias)
# output projection
self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)
# regularization
self.attn_dropout = nn.Dropout(config.dropout)
self.resid_dropout = nn.Dropout(config.dropout)
self.n_head = config.n_head
self.n_embd = config.n_embd
self.dropout = config.dropout
# flash attention make GPU go brrrrr but support is only in PyTorch nightly and still a bit scary
self.flash = (
hasattr(torch.nn.functional, "scaled_dot_product_attention") and self.dropout == 0.0
)
def forward(self, x):
B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)
# calculate query, key, values for all heads in batch and move head forward to be the batch dim
q, k, v = self.c_attn(x).split(self.n_embd, dim=2)
k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
# causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)
if self.flash:
# efficient attention using Flash Attention CUDA kernels
y = torch.nn.functional.scaled_dot_product_attention(
q, k, v, attn_mask=None, dropout_p=self.dropout, is_causal=False
)
else:
# manual implementation of attention
att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
att = F.softmax(att, dim=-1)
att = self.attn_dropout(att)
y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
y = (
y.transpose(1, 2).contiguous().view(B, T, C)
) # re-assemble all head outputs side by side
# output projection
y = self.resid_dropout(self.c_proj(y))
return y
class FineBlock(nn.Module):
def __init__(self, config):
super().__init__()
self.ln_1 = nn.LayerNorm(config.n_embd)
self.attn = NonCausalSelfAttention(config)
self.ln_2 = nn.LayerNorm(config.n_embd)
self.mlp = MLP(config)
def forward(self, x):
x = x + self.attn(self.ln_1(x))
x = x + self.mlp(self.ln_2(x))
return x
class FineGPT(GPT):
def __init__(self, config):
super().__init__(config)
del self.lm_head
self.config = config
self.n_codes_total = config.n_codes_total
self.transformer = nn.ModuleDict(
dict(
wtes=nn.ModuleList(
[
nn.Embedding(config.input_vocab_size, config.n_embd)
for _ in range(config.n_codes_total)
]
),
wpe=nn.Embedding(config.block_size, config.n_embd),
drop=nn.Dropout(config.dropout),
h=nn.ModuleList([FineBlock(config) for _ in range(config.n_layer)]),
ln_f=nn.LayerNorm(config.n_embd),
)
)
self.lm_heads = nn.ModuleList(
[
nn.Linear(config.n_embd, config.output_vocab_size, bias=False)
for _ in range(config.n_codes_given, self.n_codes_total)
]
)
for i in range(self.n_codes_total - config.n_codes_given):
self.transformer.wtes[i + 1].weight = self.lm_heads[i].weight
def forward(self, pred_idx, idx):
device = idx.device
b, t, codes = idx.size()
assert (
t <= self.config.block_size
), f"Cannot forward sequence of length {t}, block size is only {self.config.block_size}"
assert pred_idx > 0, "cannot predict 0th codebook"
assert codes == self.n_codes_total, (b, t, codes)
pos = torch.arange(0, t, dtype=torch.long, device=device).unsqueeze(0) # shape (1, t)
# forward the GPT model itself
tok_embs = [
wte(idx[:, :, i]).unsqueeze(-1) for i, wte in enumerate(self.transformer.wtes)
] # token embeddings of shape (b, t, n_embd)
tok_emb = torch.cat(tok_embs, dim=-1)
pos_emb = self.transformer.wpe(pos) # position embeddings of shape (1, t, n_embd)
x = tok_emb[:, :, :, : pred_idx + 1].sum(dim=-1)
x = self.transformer.drop(x + pos_emb)
for block in self.transformer.h:
x = block(x)
x = self.transformer.ln_f(x)
logits = self.lm_heads[pred_idx - self.config.n_codes_given](x)
return logits
def get_num_params(self, non_embedding=True):
"""
Return the number of parameters in the model.
For non-embedding count (default), the position embeddings get subtracted.
The token embeddings would too, except due to the parameter sharing these
params are actually used as weights in the final layer, so we include them.
"""
n_params = sum(p.numel() for p in self.parameters())
if non_embedding:
for wte in self.transformer.wtes:
n_params -= wte.weight.numel()
n_params -= self.transformer.wpe.weight.numel()
return n_params
@dataclass
class FineGPTConfig(GPTConfig):
n_codes_total: int = 8
n_codes_given: int = 1

39
model-card.md Normal file
View File

@ -0,0 +1,39 @@
# Model Card: Bark
This is the official codebase for running the text to audio model, from Suno.ai.
The following is additional information about the models released here.
## Model Details
Bark is a series of three transformer models that turn text into audio.
### Text to semantic tokens
- Input: text, tokenized with [BERT tokenizer from huggingface](https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertTokenizer)
- Output: semantic tokens that encode the audio to be generated
### Semantic to coarse tokens
- Input: semantic tokens
- Output: tokens from the first two codebooks of the [Encodec Codec](https://github.com/facebookresearch/encodec) from facebook
### Coarse to fine tokens
- Input: the first two codebooks from EnCodec
- Output: 8 codebooks from EnCodec
### Architecture
| Model | Parameters | Attention | Output Vocab size |
|:-------------------------:|:----------:|------------|:-----------------:|
| Text to semantic tokens | 80 M | Causal | 10,000 |
| Semantic to coarse tokens | 80 M | Causal | 2x 1,024 |
| Coarse to fine tokens | 80 M | Non-causal | 6x 1,024 |
### Release date
April 2023
## Broader Implications
We anticipate that this model's text to audio capabilities can be used to improve accessbility tools in a variety of languages.
Straightforward improvements will allow models to run faster than realtime, rendering them useful for applications such as virtual assistants.
While we hope that this release will enable users to express their creativity and build applications that are a force
for good, we acknowledge that any text to audio model has the potential for dual use. While it is not straightforward
to voice clone known people with Bark, they can still be used for nefarious purposes.

58
pyproject.toml Normal file
View File

@ -0,0 +1,58 @@
[build-system]
requires = ["setuptools"]
build-backend = "setuptools.build_meta"
[project]
name = "suno-bark"
version = "0.0.1a"
description = "Bark text to audio model"
readme = "README.md"
requires-python = ">=3.8"
authors = [
{name = "Suno Inc", email = "hello@suno.ai"},
]
# Apache 2.0
license = {file = "LICENSE"}
dependencies = [
"boto3",
"encodec",
"funcy",
"numpy",
"scipy",
"tokenizers",
"torch",
"tqdm",
"transformers",
]
[project.urls]
source = "https://github.com/suno-ai/bark"
[project.optional-dependencies]
dev = [
"bandit",
"black",
"codecov",
"flake8",
"huggingface-hub",
"hypothesis>=6.14,<7",
"isort>=5.0.0,<6",
"jupyter",
"mypy",
"nbconvert",
"nbformat",
"pydocstyle",
"pylint",
"pytest",
"pytest-cov",
]
[tool.setuptools]
packages = ["bark"]
[tool.setuptools.package-data]
bark = ["assets/prompts/*.npz"]
[tool.black]
line-length = 100

3
setup.py Normal file
View File

@ -0,0 +1,3 @@
from setuptools import setup
setup()