Linacodec: Highly compressive audio tokenizer for speech models.

Linacodec is an audio tokenizer that compresses audio into just 12.5 tokens per second (171 bps) and decodes to 48khz audio.

https://github.com/user-attachments/assets/77094afd-2d5d-490e-b283-9100e74a69a4

Key benefits

• Compression: 12.5 tokens/sec (60x more compressed than DAC).
• Audio Quality: 48khz output (much clearer then 16khz/24khz which is the standard).
• Encoder Speed: 200x realtime.
• Decoder Speed: 400x realtime(even faster with batching)
• Many Tasks: Indirectly even supports voice conversion, audio super-resolution, and audio denoising!

Why is this even useful?

Audio tokenizers directly contribute to speed, quality, and capability of TTS/ASR models. LinaCodec massively improves upon previous codecs in these areas.

• Inference Speed: Enables TTS models to run 800x realtime, 8x faster than MiraTTS!
• Fast training: High-quality TTS models can be trained in less then 1 day.
• Versatile: Works for both Text-to-Speech and Speech-to-Text unlike most other codecs.

Comparisons

Lower tokens/sec means faster models and higher sample rate means more clarity.

Usage

Simple 1 line installation:

pip install git+https://github.com/ysharma3501/LinaCodec.git

Reconstruction

from IPython.display import Audio
from linacodec.codec import LinaCodec

## load model
lina_tokenizer = LinaCodec() ## will download YatharthS/LinaCodec from huggingface

## get speech tokens and global embedding
speech_tokens, global_embedding = lina_tokenizer.encode("your_audio_path.wav")

## decode them into 48khz audio
audio = lina_tokenizer.decode(speech_tokens, global_embedding)

## display audio
display(Audio(audio.cpu(), rate=48000))

Voice conversion

## Assuming you have loaded model
source_wav = "source_wav.wav" ## the content you want
reference_wav = "reference_wav.wav" ## the timbre(style) you want

## convert voice
audio = lina_tokenizer.convert_voice(source_wav, reference_wav)

## display audio
display(Audio(audio.cpu(), rate=48000))

Audio super resolution

## get speech tokens and global embedding from 24khz wav
speech_tokens, global_embedding = lina_tokenizer.encode("your_audio_path.wav")

## decode them into 48khz audio(upsamples from 24khz-->48khz)
audio = lina_tokenizer.decode(speech_tokens, global_embedding)

## display audio
display(Audio(audio.cpu(), rate=48000))

Notes

This is heavily based of kanade-tokenizer so massive thanks to them!

The key novel parts I added are:

1. Dual-Path Vocos Decoder: Enables high-quality 48kHz reconstruction from original 24khz vocos using only 30 hours of training data (compared to the typical hundreds of hours).
2. Distilled WavLM Base+: Increased encoder speed while being similar quality.
3. Snake based upsampling: Used custom upsampling block to upscale features based off snake activation from BigVGAN.

Next steps

• Release code and model
• Release article on how kanade and Lina work so well at rates of 12.5 t/s compared to others.
• Possible paper on how these techniques can easily work on any codec.

Stars and Likes would be appreciated if found helpful, thank you.

Model link: https://huggingface.co/YatharthS/LinaCodec Email: yatharthsharma3501@gmail.com