modern-asr 0.1.6

A unified, extensible, and modern Python toolkit for LLM-based Automatic Speech Recognition (ASR).

Modern ASR

A unified, extensible, and future-proof Python toolkit for locally running state-of-the-art LLM-based Automatic Speech Recognition (ASR) models.

---

✨ Features

• 🧩 23 Models — Whisper, SenseVoice, Qwen, MiMo, FireRedASR, GLM-ASR, and more
• 🔌 Plugin Architecture — Add new models with @register_model decorator
• 🚀 Hot-Swap — Switch models at runtime without restarting
• 🌍 Multi-Language — 52 languages, 22 Chinese dialects
• 🎯 Multi-Task — Transcription, translation, diarization, emotion, events
• 💻 Local-First — All inference on-device. No APIs. No data leaves your machine.
• 🍎 Apple Silicon — MPS (Metal Performance Shaders) support on macOS
• 🐍 Modern Python — uv-native packaging, Pydantic configs, rich CLI

---

📦 Installation

From PyPI (Recommended)

pip install modern-asr

模型依赖和权重会在第一次使用时自动安装 — 你只需要输入模型名字，其余全自动化：

from modern_asr import ASRPipeline

# SenseVoice — 自动安装 funasr、modelscope，自动下载权重
pipe = ASRPipeline("sensevoice-small")

# MiMo-ASR — 自动 clone 官方仓库，自动下载 HF 权重
pipe = ASRPipeline("mimo-asr-v2.5")

# Whisper — 自动安装 openai-whisper，自动下载权重
pipe = ASRPipeline("whisper-small")

如果需要预装所有依赖（离线环境）：

pip install modern-asr[all-models]

Available extras: transformers, vllm, onnx, fireredasr, sensevoice, funasr, qwen-asr, mimo-asr, canary, glm-asr, whisper, moonshine, all-models, all-backends, all.

From Source

# Clone the repository
git clone https://github.com/vra/modern-asr.git
cd modern-asr

# Sync dependencies (recommended)
uv sync --all-extras

# Or install specific extras only
uv sync --extra transformers --extra whisper

# Or just core dependencies
uv sync

Python 3.10+ recommended. Some models (Qwen3-ASR, MiMo) require Python ≥ 3.10.

---

🚀 Quick Start

from modern_asr import ASRPipeline

# Transcribe with SenseVoice (Alibaba)
pipe = ASRPipeline("sensevoice-small")
result = pipe("audio.wav", language="zh")
print(result.text)

# Switch to Qwen3-ASR for dialect support
pipe.switch_model("qwen3-asr-0.6b")
result = pipe("audio.wav", language="zh")
print(result.text)

# English with Whisper
pipe.switch_model("whisper-small")
result = pipe("audio.wav", language="en")

---

📚 Documentation

Full documentation with Material for MkDocs:

mkdocs serve

---

🏗️ Architecture

Modern ASR is built on three layers:

1. ASRPipeline — Unified user API. Handles input normalization, task dispatch, model lifecycle.
2. ASRModel / AudioLLMModel — Adapter layer. New models often need only 8 lines of config via AudioLLMModel.
3. Backends — Transformers, vLLM, ONNX Runtime.

Adding a New Model

from modern_asr.core.audio_llm import AudioLLMModel
from modern_asr.core.registry import register_model

@register_model("my-model-1b")
class MyModel1B(AudioLLMModel):
    HF_PATH = "org/MyModel-1B"
    SUPPORTED_LANGUAGES = {"zh", "en"}
    CHUNK_DURATION = 30.0

    @property
    def model_id(self) -> str:
        return "my-model-1b"

That's it. The registry auto-discovers it at runtime.

---

🤝 Contributing

See Contributing Guide for development setup, code style, and PR checklist.

---

📄 License

Apache-2.0