Summary

GLM-ASR-Nano is an open-source automatic speech recognition (ASR) model designed to handle real-world audio complexities. It targets users needing robust transcription, especially for diverse dialects and low-volume speech, offering a competitive alternative to existing models like Whisper V3.

How It Works

This 1.5 billion parameter model employs a robust architecture optimized for challenging acoustic environments. Its design prioritizes exceptional dialect support, including Cantonese, and specialized training for low-volume or quiet speech scenarios. This approach aims to achieve state-of-the-art performance, particularly on Chinese benchmarks, by effectively capturing nuances often missed by conventional ASR systems.

Quick Start & Requirements

Installation involves pip install -r requirements.txt and sudo apt install ffmpeg. Inference can be performed using the transformers library (supporting 5.x), vLLM, or SGLang. Example inference scripts for English and Chinese audio are provided.

• Hugging Face
• ModelScope

Highlighted Details

• Achieves a state-of-the-art average error rate of 4.10 on Chinese benchmarks.
• Outperforms OpenAI Whisper V3 on multiple benchmarks while maintaining a compact 1.5B parameter size.
• Demonstrates exceptional performance on Cantonese and other dialects, as well as low-volume speech.
• Evaluated on real-world meeting scenarios (Wenet Meeting) and standard Mandarin datasets (Aishell-1).

Maintenance & Community

The project mentions a WeChat community for engagement. No specific details on core maintainers, sponsorships, or roadmap are provided in the README.

Licensing & Compatibility

The README does not explicitly state the license type or any compatibility notes for commercial use.

Limitations & Caveats

The README does not detail specific limitations, alpha status, known bugs, or unsupported platforms. The focus on specific Chinese dialects and low-volume speech might imply limitations in broader language support or standard-volume speech scenarios compared to more general-purpose models.