Modern ASR

A unified, extensible, future-proof toolkit for locally running state-of-the-art LLM-based ASR models.

简体中文 · Features · Installation · Models · Quick Start · Architecture

---

✨ Features

• 🧩 19 Models — Whisper, SenseVoice, Qwen, MiMo, FireRedASR, GLM-ASR, and more.
• 🔌 Zero-Config Plugin — Add new models via @register_model decorator.
• 🚀 Runtime Hot-Swap — Switch models without restarting the process.
• 🌍 Multi-Language — 52 languages, 22 Chinese dialects.
• 🎯 Multi-Task — Transcription, translation, diarization, emotion, events.
• 💻 Local-First — All inference on-device. No API keys. No data leaves your machine.
• 🍎 Apple Silicon — Native MPS (Metal Performance Shaders) support.
• 📦 Auto-Install — Dependencies, git repos, and HF weights are installed automatically on first use.
• 🐍 Modern Python — Pydantic configs, rich CLI, ISO-timestamped logging.

---

📦 Installation

pip install modern-asr

Dependencies and model weights are installed automatically the first time you use a model — just type its name:

from modern_asr import ASRPipeline

pipe = ASRPipeline("sensevoice-small")
pipe = ASRPipeline("mimo-asr-v2.5")
pipe = ASRPipeline("whisper-small")

For offline/air-gapped environments, pre-install everything:

pip install modern-asr[all-models]

Available extras: transformers, vllm, onnx, firered-asr, sensevoice, fun-asr, qwen-asr, mimo-asr, glm-asr, whisper, moonshine, all-models, all-backends, all.

Requirements: Python ≥ 3.10.

---

🧩 Supported Models

| Qwen3-ASR (Alibaba) | qwen3-asr-0.6b | 0.6B | 22 dialects | qwen-asr | | | qwen3-asr-1.7b | 1.7B | 22 dialects | qwen-asr |

| FunASR / Paraformer (Alibaba) | funasr-nano | 0.8B | zh/en | fun-asr | | | paraformer-zh | 0.2B | zh | fun-asr | | | paraformer-large | 0.7B | zh | fun-asr | | FireRedASR (Xiaohongshu) | fireredasr-aed | 1.1B | zh | firered-asr | | | fireredasr-llm | 8.3B | zh | firered-asr | | MiMo-ASR (Xiaomi) | mimo-asr-v2.5 | 8B | zh/dialects | mimo-asr | | MiDasheng (Xiaomi) | midashenglm-7b | 7B | audio understanding | mimo-asr |

| GLM-ASR (Zhipu AI) | glm-asr-nano-2512 | 1.5B | zh/en/yue | glm-asr | | Granite Speech (IBM) | granite-speech-3.3-8b | 8B | en | transformers | | Moonshine (Useful Sensors) | moonshine-tiny | 27M | en | moonshine |

# List all available models
python -m modern_asr list

---

🚀 Quick Start

from modern_asr import ASRPipeline

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

# Switch to Qwen3-ASR for dialects
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")
print(result.text)

---

🏗️ Architecture

Modern ASR is built on three layers:

1. ASRPipeline — Unified user API. Input normalization, task dispatch, model lifecycle.
2. ASRModel / AudioLLMModel — Adapter layer. New models often need only 8 lines of config.
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"

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

---

📚 Documentation

Full documentation with Material for MkDocs:

mkdocs serve

---

🤝 Contributing

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

---

📄 License

Apache-2.0

---