This project provides an industrial audio online policy distillation (OPD) training stack for Automatic Speech Recognition (ASR) and Text-to-Speech (TTS). It enables the distillation of compact, efficient audio models from larger, stronger teacher models, offering a data-efficient approach for developing high-performance ASR systems. The stack is designed for researchers and engineers seeking to reduce model size and training data requirements without sacrificing accuracy.

How It Works

The core ASR approach involves an autoregressive student model generating transcripts on-policy. A teacher model then scores these generated transcripts against the same audio input. The student model is updated via token-level KL divergence, calculated on the union of top-k token supports from both the teacher and student. This method leverages the teacher's knowledge on the student's current behavior, facilitating efficient distillation.

Quick Start & Requirements

• Primary install: pip install -e .
• Prerequisites: CUDA/PyTorch environment, numpy, torch, datasets, transformers, omegaconf, and verl dependencies. Specific teacher backends may require additional installations (e.g., vLLM).
• Links:
  • GitHub: https://github.com/AutoArk/open-audio-opd
  • Paper: https://arxiv.org/abs/2605.28139
  • Demo: https://huggingface.co/spaces/AutoArk-AI/Ark-ASR-0.6B
  • Model: https://huggingface.co/AutoArk-AI/ARK-ASR-0.6B

Highlighted Details

• Data Efficiency: Achieves performance comparable to Qwen3-ASR-0.6B using only 100k hours of ASR audio, a significant reduction compared to teacher model pretraining scales.
• Model Release: Offers the ARK-ASR-0.6B checkpoint, a compact ASR student model supporting multiple languages including Chinese, English, German, Japanese, French, Korean, Spanish, and others.
• Distributed Training: Employs FSDP2 for efficient distributed training, with a hostfile launcher for multi-node setups.
• Deployment: Includes a vLLM online service for ASR inference, featuring an OpenAI-style API endpoint.

Maintenance & Community

The project was released in May 2026, with recent announcements regarding its paper and online demo. A roadmap indicates planned TTS capabilities. Specific community links (e.g., Discord, Slack) or details on core contributors are not provided in the README.

Licensing & Compatibility

The license is specified as "See LICENSE" and links to a LICENSE file within the repository. The exact license type and its implications for commercial use or closed-source linking are not detailed in the README.

Limitations & Caveats

The current release focuses exclusively on ASR; TTS OPD functionality is planned but not yet implemented. Certain teacher model backends require external code (e.g., Qwen3-ASR backend) that is not included. The repository does not bundle audio files or datasets, requiring users to provide their own data and configuration paths.