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!
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
| Model | Total Tokens/Sec | Sample Rate |
|---|---|---|
| Linacodec | 12.5 | 48khz |
| DAC | 774 | 44.1khz |
| EnCodec | 300 | 24khz |
| Xcodec2 | 50 | 16khz |
| Mimi | 200 | 24khz |
Please check the repo for usage: https://github.com/ysharma3501/LinaCodec
Licence is CC-BY-4.0 meaning you can use it for any usecase(commercially/non-commercially) given you credit the original creator. Thank you.