Repository navigation
feat(speech): Pocket default agent voices + warm Pocket sessions, fp32 Kokoro fallback, Whisper small (v0.3.39) - #313
Merged
Conversation
TTS: the int8 Kokoro export sounds fuzzy and loses the top octave (~7 dB less above 7 kHz on am_michael) and on ONNX Runtime CPU it was also slower than fp32 (rtf 4.0-4.3 vs 2.3-2.6, measured on vienna). Nodes now provision kokoro-v1.0.onnx (model-files-v1.1, sha256-pinned); the worker prefers it and falls back to int8, and ensure_kokoro upgrades an int8-only node in place without ever dropping it out of TTS. STT: provision faster-whisper small instead of base, decode with beam 5 instead of greedy; both overridable per request (model, beam_size). Anthony 2026-11-11: "it still has static ... not crisp and clean like it should be ... can we get better quality out of these tts and stt models" Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…effort Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…et sessions Anthony 2026-11-11: "its still computerish. we need to be more human ... use that for default voices everywhere". - speech_profiles: every agent's engine/voice is now Pocket (its existing pocket_voice); the old Kokoro/Piper voice moves to fallback_engine / fallback_voice, exposed on /v1/voices and in `c0mpute tts voices`. - Warm sessions serve Pocket: the session voice picks the engine (voice ids are unique across engines); the worker's serve loop loads Pocket once and caches each voice's prompt state. Warm Pocket ran ~1x realtime on 8 threads on a busy vienna (fp32 Kokoro 2.3x there). - One warm worker per engine, shared by every session on it: a five-voice show holds one model in RAM, not five (MAX_SESSIONS 4 -> 16; it used to refuse the fifth voice). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ralyodio
changed the title
ThreatCrush Security Scan14 finding(s) HIGH/CRITICAL: 1 | MEDIUM: 9 | LOW: 4
Snippets are redacted; ThreatCrush never prints matched credential material. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
TTS sounded fuzzy and not crisp on TUI Weekly (c0mpute voices). Cause: we serve the int8 Kokoro export.
Side by side on vienna, same text, kokoro-onnx 0.6.1:
int8 is both duller and slower on ORT CPU. This PR:
kokoro-v1.0.onnx(model-files-v1.1, sha256 beb0d184…), worker prefers it, int8 kept as fallbackensure_kokoroupgrades int8-only nodes in place; failure keeps serving int8beam_size5 (was 1), both per-request overridable🤖 Generated with Claude Code