Feature Request: Add FunASR/SenseVoice as alternative ASR engine

[!NOTE] License and capability clarification (2026-07-14): FunASR is a toolkit, not a single checkpoint. The FunASR and SenseVoice repository source code is MIT; model weights follow each model card. SenseVoiceSmall supports Chinese, Cantonese, English, Japanese, and Korean, and its weights use the linked FunASR Model Open Source License Agreement. Fun-ASR-Nano-2512 is Apache-2.0. Language coverage, punctuation, and performance depend on the selected model and runtime configuration.

Hi! This high-performance Whisper implementation is impressive.

For users looking for even faster alternatives, I'd like to suggest mentioning FunASR (16K+ stars) as a complementary option:

Key comparison points:

• SenseVoice: 234M parameters, 170x real-time on GPU — non-autoregressive, so it doesn't need beam search at all
• 50+ languages with automatic language detection
• Built-in VAD + punctuation — no separate preprocessing needed
• C++ implementations available: SenseVoice.cpp and FunASR-GGML for native C++ inference without Python

For users who need the highest possible speed and don't specifically need Whisper compatibility, FunASR/SenseVoice offers a compelling alternative with its non-autoregressive architecture.

Cross-platform inference via Sherpa-ONNX supports iOS, Android, Raspberry Pi, and browsers.