Feature: FunASR as alternative ASR engine (faster + multilingual + runs on CPU)

[LauraGPT]

Context

RTranslator does real-time on-device translation, which requires fast speech recognition. FunASR (16.5K stars, MIT) provides models that are significantly faster than Whisper, especially on CPU — which matters for mobile devices.

Key advantages for mobile/on-device use

	Whisper	FunASR SenseVoice	FunASR Paraformer
CPU speed	Slow (~1x realtime)	17x realtime	15x realtime
Model size	1.5GB (large)	234MB	220MB
Languages	57	50+	zh/en
Architecture	Autoregressive	Non-autoregressive	Non-autoregressive
Latency per 5s	~5s on CPU	<300ms on CPU	<400ms on CPU

The non-autoregressive architecture is the key difference — it processes the entire audio segment in a single forward pass, making it inherently faster and more suitable for real-time applications on resource-constrained devices.

Model options for mobile

• SenseVoice-Small (234MB): 50+ languages, best for general multilingual use
• Paraformer-zh (220MB): Best Chinese + English accuracy

Both models have ONNX exports available for mobile deployment:

• ONNX Runtime Mobile compatible
• funasr-onnx package for Python
• C++ runtime available

Resources

• ONNX models: https://github.com/modelscope/FunASR/tree/main/runtime/python/onnxruntime
• C++ runtime: https://github.com/modelscope/FunASR/tree/main/runtime
• Android deployment guide available in docs

Happy to help with integration details if interested.