Feature: Add FunASR/SenseVoice as local STT engine

Motivation

agenticSeek currently uses Vosk for fully local, real-time microphone transcription. FunASR is a local speech toolkit, and iic/SenseVoiceSmall could be useful as an optional utterance-level backend when its language and accuracy profile fits the user.

This should be additive rather than a direct replacement. The existing Vosk path supports streaming partials, grammar constraints, word timestamps, and TTS echo filtering. SenseVoiceSmall normally consumes a completed utterance, so an integration must preserve the current real-time behavior or make the trade-off explicit.

Proposed scope

• Keep Vosk as the default streaming backend.
• Add a configurable SenseVoice backend, or use SenseVoice as an optional second pass after Vosk detects an utterance boundary.
• Preserve the existing microphone lifecycle and TTS echo-filter behavior.
• Load FunASR lazily and show an actionable dependency/model error when the optional backend is selected.
• Support CPU inference; treat any accelerator support as an optional configuration.
• Document model download size, measured memory usage, supported languages, and the model-specific license.
• Add focused tests for backend selection, missing optional dependencies, final-result normalization, and echo-filter integration.

Capabilities and constraints

• FunASR toolkit source and the SenseVoice repository source are MIT-licensed; model weights follow their individual model cards.
• SenseVoiceSmall supports Chinese, Cantonese, English, Japanese, and Korean. Language coverage and performance vary by checkpoint.
• SenseVoiceSmall can emit language, emotion, and acoustic-event tags; callers should decide whether to retain or strip them.
• A two-clip CPU benchmark against agenticSeek's current Vosk model is included below. It is evidence for this integration discussion, not a universal quality or speed claim.

References

• FunASR: https://github.com/modelscope/FunASR
• SenseVoice: https://github.com/FunAudioLLM/SenseVoice
• SenseVoiceSmall model card: https://huggingface.co/FunAudioLLM/SenseVoiceSmall
• Current agenticSeek STT implementation: https://github.com/Fosowl/agenticSeek/blob/main/sources/speech_to_text.py