Codec BPE is an implementation of Acoustic BPE (Shen et al., 2024), extended for RVQ-based Neural Audio Codecs such as EnCodec (Défossez et al., 2022), DAC (Kumar et al., 2023), and Mimi (Défossez et al., 2024). Built on top of the HuggingFace Tokenizers library.

Codec BPE flattens multi-level codes from Residual Vector Quantizers (RVQ) and converts them into unicode strings for tokenization into compressed token sequences. For example, a single Codec BPE token might represent a 4-gram of codes from 4 codebooks representing a single acoustic unit, a 6-gram comprising a whole acoustic unit and half of the next one, or even an 8-gram represnting two whole acoustic units. Depending on the codec, vocab size and type of audio, this can yield savings of 2-5x in sequence length compared to directly modeling the flattened codebooks.

Using Codec BPE allows efficient audio language modeling with multi-level codecs to be done with vanilla LLM architectures, meaning no custom architecture is needed to deal with modeling the RVQ. Your model will already be compatible with the full ecosystem of training and inference tools available for HuggingFace Transformers, such as vLLM!

2024-09-20

• Added support for Kyutai Lab's Mimi codec, an amazing new codec with a 12.5 Hz framerate! Simply add --use_mimi when encoding audio with codec_bpe.audio_to_codes to encode using the Mimi model. More info here.

  Note: Until Mimi is included in a stable release of HuggingFace Transformers, you need to install Transformers from source:

  pip install git+https://github.com/huggingface/transformers.git@main

2024-09-19

• Initial Release!

pip install codec-bpe

Use your codec of choice (e.g., EnCodec, DAC, Mimi) to encode your audio into a torch tensor or numpy array of codes of shape (num_codebooks, length), then use the provided converter methods to convert to and from unicode strings.

Note: In the Acoustic BPE paper, a single-level codec was used (HuBERT + k-means), where each encoded timestep consisted of a single code which was converted to a single unicode character. Here, we support multi-level codecs based on Residual Vector Quantizers. If num_codebooks > 1, a flattening pattern is used to interleave all codebooks into a single level before mapping to unicode. For example, if 4 codebooks are used then each encoded timestep would consist of 4 codes (one from each codebook) and would be converted to a unicode 4-gram.

Example: audio language modeling using EnCodec 24 kHz at 3 kbps (4 codebooks):

import torch
import librosa
import soundfile as sf
from transformers import (
    EncodecModel, 
    AutoModelForCausalLM,
    AutoProcessor, 
    AutoTokenizer,
)
from codec_bpe import codes_to_chars, chars_to_codes

# load a Codec BPE tokenizer and compatible language model
device = "cuda" if torch.cuda.is_available() else "cpu"
tokenizer = AutoTokenizer.from_pretrained("output/my_tokenizer")
model = AutoModelForCausalLM.from_pretrained("output/my_model").to(device)

# load the EnCodec model
encodec_modelname = "facebook/encodec_24khz"
encodec_model = EncodecModel.from_pretrained(encodec_modelname).to(device)
encodec_processor = AutoProcessor.from_pretrained(encodec_modelname)

# (1) encode audio using EnCodec
audio, sr = librosa.load("some_audio.mp3", sr=encodec_model.config.sampling_rate, mono=True)
inputs = encodec_processor(raw_audio=audio, sampling_rate=sr, return_tensors="pt").to(device)
with torch.no_grad():
    encoded_audio = encodec_model.encode(**inputs, bandwidth=3.0).audio_codes[0, 0]

# (2) convert the audio codes to a unicode string and tokenize it
unicode_str = codes_to_chars(encoded_audio, codebook_size=encodec_model.config.codebook_size)
inputs = tokenizer(unicode_str, return_tensors="pt").to(device)

# (3) generate tokens from the model
outputs = model.generate(**inputs, do_sample=True, max_new_tokens=300)

# (4) detokenize the output back into a unicode string and convert it back to audio codes
unicode_str_2 = tokenizer.decode(outputs[0], skip_special_tokens=False)
encoded_audio_2 = chars_to_codes(
    unicode_str_2, 
    num_codebooks=encoded_audio.shape[0], 
    codebook_size=encodec_model.config.codebook_size, 
    return_tensors="pt",
).to(device)

# (5) decode the generated audio using EnCodec
with torch.no_grad():
    audio_2 = encodec_model.decode(encoded_audio_2.unsqueeze(0).unsqueeze(0), [None]).audio_values[0, 0]
sf.write("some_audio_output.wav", audio_2.cpu().numpy(), sr)

To train a tokenizer from audio files:

1. Use your codec of choice (e.g., EnCodec, DAC, Mimi) to encode each audio file into a directory of numpy arrays (.npy files):

   # encode audio files using EnCodec 24 kHz at 3 kbps (4 codebooks)
   python -m codec_bpe.audio_to_codes \
       --audio_path path/to/audio \
       --encodec_model facebook/encodec_24khz \
       --bandwidth 3.0
   
   # encode audio files using first 4 codebooks of DAC 44kHz
   python -m codec_bpe.audio_to_codes \
       --audio_path path/to/audio \
       --dac_model 44khz \
       --n_quantizers 4 \
       --use_dac
   
   # encode audio files using first 6 codebooks of Mimi (24kHz)
   python -m codec_bpe.audio_to_codes \
       --audio_path path/to/audio \
       --mimi_model kyutai/mimi \
       --n_quantizers 6 \
       --use_mimi

2. Suppose you want to use the first 4 codebooks of EnCodec 24 kHz, run:

   python -m codec_bpe.train_tokenizer \
       --codes_path output/codes/encodec_24khz/mono \
       --num_codebooks 4 \
       --codebook_size 1024 \
       --codec_framerate 75 \
       --chunk_size_secs 30 \
       --vocab_size 30000 \
       --pad_token "<pad>" \
       --save_path output/my_tokenizer

   Here:

   • num_codebooks specifies how many codebooks should be used (in a flattened pattern) when converting each timestep to unicode. For example, EnCodec 24kHz uses 2 codebooks at 1.5 kbps, 4 codebooks at 3 kbps, 8 codebooks at 6 kbps, etc. Note: when encoding the audio files, you should use at least as many codebooks as you plan to specify here.
   • codebook_size specifies the size of the codebook. EnCodec 24 kHz uses a codebook size of 1024.
   • codec_framerate specifies the framerate (number of timesteps per second) of the codec. EnCodec 24 kHz generates 75 timesteps per second.
   • chunk_size_secs specifies the number of timesteps (in seconds) that get converted to unicode and returned to the underlying Tokenizers trainer at a time.
   • vocab_size specifies the number of tokens (including the base vocabulary of individual unicode characters) that you want your tokenizer to have. The base vocabulary size is num_codebooks x codebook_size. For example, the command above would yield a tokenizer with a base vocabulary of 4096 individual unicode character tokens, each representing a single code from a single codebook, and 25,904 merged "ngram" tokens.

   See train_tokenizer.py for a complete list of supported arguments.

You may want to train a new Codec BPE tokenizer and then export its trained vocabulary to an existing Transformers tokenizer. For example, extending the Llama3, Mistral, Qwen, etc. tokenizers for multimodal text-audio language modeling.

Suppose you have trained your Codec BPE tokenizer and saved it to output/tokenizer.json and you want to extend the Mistral-7B-v0.1 tokenizer with its vocabulary, run:

python -m codec_bpe.extend_tokenizer \
    --existing_tokenizer mistralai/Mistral-7B-v0.1 \
    --codec_bpe_tokenizer output/my_tokenizer \
    --audio_start_token "<audio>" \ # optional
    --audio_end_token "</audio>"    # optional

This will simply add every token in output/tokenizer.json to the mistralai/Mistral-7B-v0.1 tokenizer as a special token and save a copy of the latter.

If the added Codec BPE unicode tokens would conflict with existing tokens in the vocabulary, there are two options to mitigate this:

1. Override the default unicode offset using the unicode_offset argument for both codec_bpe.train_tokenizer and codec_bpe.extend_tokenizer. By default, unicode characters from the CJK Unified Ideographs block are used, following the Acoustic BPE paper. You can set unicode_offset to a different value to use a different unicode block that doesn't conflict with your existing vocabulary.

2. Use the use_special_token_format argument for codec_bpe.extend_tokenizer. This wraps each unicode character in each ngram with <>. For example, the 4-gram token "一刁嘂娃" would be converted to a token containing the string "<一><刁><嘂><娃>". This format is more verbose, but should virtually eliminate the possibility of a vocabulary conflict:

   python -m codec_bpe.extend_tokenizer \
       --existing_tokenizer mistralai/Mistral-7B-v0.1 \
       --codec_bpe_tokenizer output/my_tokenizer \
       --audio_start_token "<audio>" \ # optional
       --audio_end_token "</audio>" \  # optional
       --use_special_token_format

   Then when preparing audio for tokenization with the extended tokenizer, you can pass the same argument to the codes_to_chars function:

   # convert codes to unicode string
   unicode_str = codes_to_chars(encoded_audio, codebook_size=model.config.codebook_size, use_special_token_format=True)

   It is unnecessary to pass this argument to chars_to_codes - it will automatically detect and remove the special token format before converting back to codes.