# jishengpeng/WavTokenizer

[ICLR 2025] SOTA discrete acoustic codec models with 40/75 tokens per second for audio language modeling

Repository: https://github.com/jishengpeng/WavTokenizer
Canonical: https://ross.abutalabs.com/products/wavtokenizer
Language: Python
License: MIT
License Family: permissive
Topics: acoustic, audio-representation, codec, gpt4o, music-representation-learning, semantic, speech-representation, text-to-speech, speech-language-model, dac, encodec, soundstream
Last push: 2025-03-02T03:53:58+00:00

## Health v2 (maintenance only)
Score: 27/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 9, release rhythm 35, longevity 52
- inputs: {"age_days": 734, "days_push": 549, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1316, forks 115 (observed 2026-08-28T04:04:20.898970+00:00)

## What it is
WavTokenizer is a state-of-the-art discrete neural audio codec that compresses speech, music, and general audio into only 40 or 75 discrete tokens per second. It is designed as an acoustic tokenizer for audio language models (like GPT-4o-style models) and text-to-speech pipelines, with pretrained checkpoints on Hugging Face.

## Use cases
- tokenize speech and audio into discrete codes for audio language models
- reconstruct audio from raw wav files with high fidelity
- build text-to-speech systems with a compact acoustic tokenizer
- replace EnCodec, DAC, or SoundStream with fewer tokens per second
- extract semantic audio representations for speech and music
- train speech language models on discrete audio tokens

## When to choose
- you need an extremely low frame rate (40/75 tokens per second) discrete audio representation
- you are building audio language models or TTS systems and need semantic-rich tokens
- you want strong audio reconstruction quality at low bitrates
- you want a drop-in alternative to EnCodec or DAC

## When to avoid
- you need real-time streaming codec operation
- you need production-grade audio compression for storage rather than ML tokenization
- you work outside the PyTorch/Python ecosystem

## Facets
- artifact type: library
- maturity: active
- function: audio-processing, machine-learning, serialization, speech-recognition, llm-training
- domain: speech-processing, machine-learning, large-language-models
- platform: python, cross-platform
- tags: neural-codec, audio-tokenizer, text-to-speech, discrete-tokens, iclr-2025, pytorch, audio, natural-language-processing

## Member repositories
- jishengpeng/WavTokenizer (main) score 27

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:20.898970+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T04:48:29.552961+00:00, confidence not recorded.
  - readme: https://github.com/jishengpeng/WavTokenizer (fetched 2026-08-28T04:04:20.898970+00:00, sha a00befb18894)
- Data as of 2026-08-30T08:39:29.467469+00:00.