Feature Request: Add SenseVoice/Paraformer as STT model option

[LauraGPT]

Feature Request

Kalosm provides instant, controllable local AI models in Rust. Suggesting SenseVoice as a STT model option alongside Whisper.

Why SenseVoice for Kalosm?

• Non-autoregressive — simpler inference implementation (no beam search/autoregressive loop)
• 5x faster than Whisper — better for real-time/interactive use
• 234M params — lighter weight
• 50+ languages — good multilingual coverage
• ONNX available — models exportable to ONNX for Rust integration via ort/tract
• GGML available — FunASR-GGML provides GGUF format models

Rust integration paths

1. ONNX Runtime (via ort crate): SenseVoice ONNX models available
2. GGML (via ggml-rs): FunASR-GGML provides GGUF models
3. C bindings: Sherpa-ONNX has a C API that could be wrapped

The non-autoregressive architecture means the Rust implementation would be simpler than Whisper — a single forward pass produces all tokens without needing an autoregressive decoding loop.

Resources

• FunASR-GGML: https://github.com/huaxin0/FunASR-GGML
• SenseVoice.cpp: https://github.com/lovemefan/SenseVoice.cpp
• FunASR: https://github.com/modelscope/FunASR (16.6K stars)
• SenseVoice: https://github.com/FunAudioLLM/SenseVoice (8.3K stars)