feat: local STT integration via cf_voice.io (whisper.cpp) #4

New issue

Closed

opened 2026-04-06 13:05:56 -07:00 by pyr0ball · 1 comment

pyr0ball commented

2026-04-06 13:05:56 -07:00

Owner

Wire whisper.cpp through cf_voice.io as the Free-tier speech-to-text backend.

Requirements:

VAD (Voice Activity Detection) gates STT to suppress silence segments
Noise gate filters low-energy frames
Ring buffer size configurable via env (default: 30s)
Model size configurable (tiny/base/small); default: base for local
Latency target: annotation visible within 1.5s of utterance end on Heimdall hardware

Tier gate: local whisper.cpp = Free. Cloud STT (faster, accent-robust) = Paid.

Blocked on: cf_voice.io being available in circuitforge-core.

Wire whisper.cpp through `cf_voice.io` as the Free-tier speech-to-text backend. Requirements: - VAD (Voice Activity Detection) gates STT to suppress silence segments - Noise gate filters low-energy frames - Ring buffer size configurable via env (default: 30s) - Model size configurable (tiny/base/small); default: base for local - Latency target: annotation visible within 1.5s of utterance end on Heimdall hardware Tier gate: local whisper.cpp = Free. Cloud STT (faster, accent-robust) = Paid. Blocked on: cf_voice.io being available in circuitforge-core.

pyr0ball added this to the Notation — v0.1.x milestone 2026-04-06 13:05:56 -07:00

pyr0ball added the

enhancement

backlog

cf-core-dep

labels 2026-04-06 13:05:56 -07:00

pyr0ball referenced this issue

2026-04-06 13:11:42 -07:00

feat: Elcor mode easter egg discovery path #15

pyr0ball referenced this issue

2026-04-06 13:43:21 -07:00

feat: cf_voice dependency — Circuit-Forge/circuitforge-core#39 #3

pyr0ball closed this issue

2026-04-11 09:50:24 -07:00

pyr0ball commented

2026-04-11 09:50:24 -07:00

Author

Owner

Done. cf_voice.stt.WhisperSTT wraps faster-whisper with async thread pool execution and a 50-word rolling session prompt for cross-chunk context continuity. Lazy-loaded on first real classify call, gated by CF_VOICE_STT=1 (default on). Confirmed working.

Done. `cf_voice.stt.WhisperSTT` wraps faster-whisper with async thread pool execution and a 50-word rolling session prompt for cross-chunk context continuity. Lazy-loaded on first real classify call, gated by `CF_VOICE_STT=1` (default on). Confirmed working.