# AutoArk/open-audio-opd

Industrial audio online policy distillation (OPD) training stack for ASR and TTS, distilling compact audio models from stronger   teacher models.

Repository: https://github.com/AutoArk/open-audio-opd
Canonical: https://ross.abutalabs.com/products/open-audio-opd
Language: Python
License: NOASSERTION
License Family: other
Last push: 2026-06-05T13:46:56+00:00

## Health v2 (maintenance only)
Score: 52/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 86, release rhythm 35, longevity 7
- inputs: {"age_days": 100, "days_push": 89, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases, young, no_license
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1007, forks 62 (observed 2026-08-28T04:03:12.313389+00:00)

## What it is
An industrial training stack for online policy distillation (OPD) of audio models, distilling compact ASR (and planned TTS) student models from stronger teacher models using token-level KL on top-k support. It builds on THUNLP/OPD and a vendored copy of verl, with FSDP2 distributed training support.

## Use cases
- distill a compact ASR model from a larger teacher model
- run online policy distillation training for speech recognition
- train audio models with FSDP2 distributed training
- reproduce the ARK-ASR-0.6B training recipe
- apply teacher-scored token-level KL training to speech models
- set up an on-policy rollout training pipeline for audio models

## When to choose
- you need to distill a small ASR model from a stronger teacher with on-policy rollouts
- you want a production-style audio distillation stack with FSDP2 distributed training
- you're researching online policy distillation for speech recognition

## When to avoid
- you need a ready-to-use inference or ASR API rather than a training stack
- you need TTS distillation today - it's still on the roadmap
- you want a framework that ships with datasets or audio files - all data paths are user-supplied

## Facets
- artifact type: library
- maturity: active
- function: llm-training, speech-recognition, tts, machine-learning, gpu-computing
- domain: speech-processing, machine-learning, deep-learning
- platform: python
- tags: knowledge-distillation, online-policy-distillation, asr, tts, fsdp2, distributed-training, verl, training-stack, audio, gpu, linux

## Member repositories
- AutoArk/open-audio-opd (main) score 52

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:03:12.313389+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 2026-08-30T07:12:23.212131+00:00, confidence not recorded.
  - readme: https://github.com/AutoArk/open-audio-opd (fetched 2026-08-28T04:03:12.313389+00:00, sha a82aaea69f36)
- Data as of 2026-08-30T08:39:29.467469+00:00.