cc-voice: implement STT via Moonshine (/listen)

[qte77]

Summary

Complete the planned STT integration in cc-voice using Moonshine for local speech-to-text input via /listen.

Reasoning

• Already listed as planned in cc-voice plugin (v0.2.0) — /listen stub exists
• No other community plugin provides both TTS + STT
• Growing voice-mode interest in CC community (voice dictation listed as hot feature in claude-code-best-practice)
• Local-first via Moonshine aligns with privacy-first design (no cloud STT dependency)

Key features to consider

• /listen command to start mic capture
• /listen --stop to end and inject transcript as user input
• Moonshine model selection (tiny/base/small for latency/accuracy tradeoff)
• PTY proxy integration matching existing /speak architecture

References

• https://github.com/shanraisshan/claude-code-best-practice — voice dictation as hot feature
• https://github.com/usefulmachines/moonshine — local STT model

[m13v]

local STT for Claude Code would be a great addition. from building on-device transcription with WhisperKit (CoreML-based, similar use case to Moonshine), a few things that will matter for the UX:

model loading time is the hidden latency killer. the first invocation of /listen will feel slow because the model needs to load into memory. lazy-loading on first use and keeping the model resident across invocations is important. on Apple Silicon, WhisperKit base loads in ~2s, Moonshine should be faster given its smaller footprint.

for the PTY integration, the key question is whether you buffer the full utterance and inject it as one block, or stream interim results. streaming looks cool but causes issues if the user sees partial text and starts editing before the final result replaces it. we went with buffer-then-inject and it felt cleaner.

one more thing: make sure to downsample to 16kHz mono before feeding the model. macOS mic defaults to 48kHz stereo and passing that directly tanks accuracy on every Whisper-family model.

[m13v]

our on-device transcription service with model management, audio preprocessing, and the buffer approach for injecting final transcripts: https://github.com/m13v/fazm/blob/main/Desktop/Sources/TranscriptionService.swift

it is WhisperKit/CoreML-specific but the model loading, 16kHz conversion, and chunk sizing patterns apply to Moonshine too.