User reported the 1.5x base made Korean pronunciation mushy/slurred. Dial the
default WSAI_TTS_SPEED back to 1.2: still noticeably faster than the original
1.0, but clear. Emotion multipliers scale off base as before.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
User reported the emotion samples sounded the same and the speed change wasn't
noticeable. Two changes:
- Add a "base" (기본/neutral) emotion at speed 1.0x so the plain base voice can
be selected explicitly via a [기본] tag and auditioned against the others.
- Bump the WSAI_TTS_SPEED default 1.15 -> 1.5 for a clearly faster base voice.
Emotion multipliers scale off base, so every emotion speeds up together.
Also extends gen_emotion_samples.py to emit one wav per emotion (incl. 기본)
plus a stitched all-in-one, so each emotion can be delivered as a separate clip.
Verified: 29 tests pass; match_emotion('기본') == 'base'; per-emotion synthesis
at base 1.5 produces distinct clip durations (base 4.9s vs happy 3.3s).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
User asked for a bit faster speech. Pitch shift is already disabled, so the
earlier "monster voice" risk from stacking pitch on a fast base is gone — a
modest 1.15x base is natural. Emotion multipliers scale off base, so every
emotion gets the bump too. Also adds tests/gen_emotion_samples.py, a dev
utility that synthesizes one clip per canonical emotion (announce name at
neutral speed, then a sample sentence steered by that emotion's tag) and
stitches them into a single wav for auditioning the emotion palette.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The emotional TTS path applied a librosa post-hoc pitch_shift (±1–3 semitones)
on top of a 1.3x-fast Melo base, producing a robotic "monster" delivery. Zero
out the pitch column for every emotion so pitch_shift is never invoked (the
worker's _pitch_shift already no-ops on 0.0), and drop the default synthesis
speed to 1.0. Emotion is now conveyed by speed alone — natural, artefact-free.
The pitch column is retained so a proper pitch method can be re-enabled later.
Verified: 29 tests pass; real 2-emotion synthesis on CUDA yields a clean wav
with pitch=0.0 on all segments.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The STT/TTS worker _ensure() treated a spawned-but-not-yet-handshaked
subprocess as ready, so a voice turn arriving during warmup read the same
stdout StreamReader concurrently with the warmup handshake and crashed with
"readuntil() called while another coroutine is already waiting for incoming
data". Add a _start_lock + _ready flag so (re)start and the ready handshake
run atomically and callers wait for real readiness before reading stdout.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Emotion tags now steer delivery rather than being read aloud. parse_segments()
splits a reply on [감정] tags: a recognised emotion word switches the pitch and
speed of the text that follows (and is dropped), while a non-emotion bracket
(e.g. [1번]) keeps its inner words as spoken content. Emotions can change
mid-reply, so a single turn is synthesised as several pitch-shifted segments and
concatenated in the melo worker (librosa pitch_shift, warmed at startup).
The emotion vocabulary is grounded in Azure Neural TTS speaking styles plus
Ekman's basic emotions, with Korean synonyms. The brain persona is updated to
emit inline tags from that set.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Claude replies with markdown/backticks by default; MeloTTS's Korean text
normaliser has no entry for '`' and dies with KeyError: '`', so any reply
mentioning a command/code block crashed the whole voice turn (500 on
/api/voice-turn). Fix at the shared synth() choke point with
normalize_for_speech(), which flattens code fences/inline code/links/markdown
and guarantees no backtick reaches the worker — covering both the dashboard
voice turn and the Discord speak() bridge. Also add a PERSONA line asking the
model to avoid markdown (belt-and-suspenders; the code strip is the real fix).
errors_total never moved for turn-level failures: it was only bumped by
log("error") events, and the dashboard voice path calls turn.finish(error=...)
without logging. Emit one error-level log event from Turn.finish() when a turn
ends in error, so both the server counter and the browser SSE mirror stay
consistent, guarded to count at most once. Drop the now-redundant pipeline
log("error") to avoid double counting and remove the dead _publish stub.
Verified: raw backtick -> worker KeyError '`' reproduced; after fix real
MeloTTS synth of a backtick+fenced reply succeeds; /api/voice-turn returns 200
with a wav body on a backtick reply and errors_total stays 0, and an induced
synth failure returns 500 with errors_total incrementing to exactly 1. Full
suite 18 passed.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The bot (dave/bot.mjs) previously only joined the channel and counted audio
frames — it never fed STT or spoke back. Wire the real loop:
- Node bot: buffer each speaker's Opus->PCM utterance until AfterSilence,
wrap as WAV, POST to the Python voice-turn endpoint, then play the returned
reply wav into the channel via an AudioPlayer (ffmpeg->Opus). Skips its own
audio, dedupes overlapping subscriptions, and ignores sub-0.35s noise.
- Python: new `python -m wsai --voice-server` serves /api/voice-turn — decode
the uploaded utterance, GPU faster-whisper STT, produce a reply (echo of what
was heard for now), GPU MeloTTS synth, return the reply wav (recognised/reply
text ride along as X-Heard/X-Reply headers). Both engines pre-warmed; turns
show in the dashboard feed. MeloTTS.synth() extracted for direct wav reuse.
Echo mode verifies listening+speaking+GPU recognition entirely in Discord; the
Claude brain is the next slice. Verified the endpoint round-trip: utterance wav
-> correct Korean X-Heard/X-Reply + a WAVE reply on device=cuda. 12 tests pass,
node --check clean.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Both voice backends defaulted to CPU. Fix the "CUDA unavailable" gaps so
everything that benefits from the RTX 5050 uses it:
- MeloTTS venv had CPU-only torch (2.12.0+cpu) -> installed Blackwell-capable
torch/torchaudio 2.11.0+cu128 (sm_120 verified with a real GPU matmul).
- faster-whisper CUDA loaded but transcribe() died with "libcublas.so.12 not
found": installed nvidia-cublas-cu12 + nvidia-cudnn-cu12 into the whisper
venv and inject those nvidia/*/lib dirs into the worker's LD_LIBRARY_PATH at
spawn (the loader only honours it at exec).
- WSAI_WHISPER_DEVICE / WSAI_MELO_DEVICE now default to "auto": pick CUDA when
present, else CPU, and each worker falls back to CPU if a CUDA load fails so
the voice loop never dies on a GPU-less host.
Verified end-to-end through the real backend classes: both workers report
"ready on cuda"; steady-state STT ~170ms (was ~1350ms CPU), TTS ~4s first call
vs ~23s CPU. All 12 tests pass.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Step 3 (귀): add WhisperSTT + whisper_worker, a warm out-of-venv worker
mirroring the MeloTTS shape (whisper312 venv, small/int8 on CPU). transcribe()
closes the voice round trip (MeloTTS wav -> whisper text); utterances() turns an
injected audio_source into Utterances (Discord voice feed pending). Wired into
factory as WSAI_STT=whisper.
Also address the arbiter's TTS follow-ups:
- melo worker error handling: capture stderr (drained in a bounded background
task so the pipe can't fill), surface the real failure cause, and defend
against an empty/invalid ready line instead of dying on JSONDecodeError.
- pipeline pre-warm: load slow backends (warmup()) at startup so the first
utterance is answered warm; a warmup failure is logged, not fatal.
Verified: real TTS->STT round trip recovers the sentence near-perfectly;
warm transcribe ~1.2s (CPU). 12 tests pass.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Adds a MeloTTS backend that runs the model in its own melo311 interpreter
as a long-lived worker (melo_worker.py), loaded once and fed synthesis
requests over a stdin/stdout JSON protocol. fd1 is split from fd2 in the
worker so MeloTTS's stdout progress chatter can't corrupt the protocol.
Each speak() writes a wav and hands the path to a pluggable sink (the
Discord voice step will swap in "play into the call"). factory wires
tts=melo; pipeline.aclose now also tears down the tts worker.
Verified (CPU): model load ~7.9s once, then a short reply synthesizes in
~0.86s (within the ~1s budget); wav is valid 44.1kHz PCM. GPU (cuda) is
selectable via WSAI_MELO_DEVICE for lower latency, pending GPU approval.
7 smoke tests still pass.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
