Sentence text in typing bubble:
- Backend sends 'text' field with each audio SSE event (the sentence
being synthesised)
- Frontend audio queue stores {url, text} pairs
- playNextAudio() writes the sentence text into the '...' typing bubble
when LLM tokens haven't arrived yet (convCurrentSentenceBubble)
- convCurrentSentenceBubble cleared as soon as first LLM token arrives
so normal streaming takes over seamlessly
VAD noise fixes:
- VAD_THRESHOLD: 0.01 → 0.02 (background noise no longer counts as speech)
- VAD_MIN_REC_MS: 400 → 800ms (8/10s wait before silence detection starts,
gives user time to begin speaking without initial noise triggering send)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Speech gate (silence detection):
- vadHadSpeech flag: VAD auto-stop cancels without calling STT when no
speech was detected (fixes "[STT] No speech detected → gibberish" loop)
- cancelNextBlob flag: onstop skips processBlob when VAD cancels silently
- vadLastVoiceMs: gates preview transcription on actual detected speech
(prevents "reich" hallucination on initial silence chunks)
Hallucination filter:
- Client: HALLUCINATION_RE strips known Whisper phantoms from preview
- Server: _is_hallucination() in generate() treats "reich" / "danke" /
"thank you" etc. as "No speech detected" → never reaches LLM
Latency:
- VAD_SILENCE_MS: 1500 → 1000 ms (sends 500 ms sooner per turn)
- VAD_MIN_REC_MS: 500 → 400 ms
- MediaRecorder timeslice: 2500 → 1500 ms (preview text updates faster)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replace serial LLM-wait-TTS with overlapped execution:
- LLM streams via background thread → asyncio.Queue (non-blocking event loop)
- _sentence_split() detects sentence boundaries in the token stream
- asyncio.create_task fires TTS for each sentence immediately — TTS for
sentence 1 runs while LLM is still generating sentences 2, 3, …
- Audio chunks stream to frontend in order as each task completes
- Time-to-first-audio drops from (LLM total + TTS total) to
roughly (LLM time-to-first-sentence + TTS latency for one sentence)
Frontend audio queue:
- enqueueAudio() / playNextAudio() chain multi-chunk responses seamlessly
- clearAudio() stops playback and cancels queue on new turn or mic click
- scheduleAutoMic() waits for queue to drain before restarting mic
- Error paths clear the queue to avoid stale audio playing after failure
Also fix missing contextlib import (silent bug when audio temp files
needed cleanup in the STT path).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Perceived startup speed:
- index.html: animated shimmer skeleton (header + toolbar + 8 voice cards)
visible immediately; fades out when loader.js finishes
- WaveSurfer (57KB) and MDI icon font (394KB woff2) now served from
static/vendor/ — removes 3 render-blocking external requests from <head>
- Flag-icons CSS loaded async (rel=preload onload trick) — non-blocking
loader.js:
- Fade out skeleton + remove from DOM (300ms transition)
- Reveal page-sections after JS finishes loading
Conversation playground:
- Qwen3 thinking mode fix: fall back to delta.reasoning_content when
delta.content is empty so think-only LLM turns produce visible output
- Better error message with /no-think hint when LLM returns empty
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
UploadFile | None = File(None) with from __future__ import annotations
caused FastAPI to still treat audio as required when omitted.
Changed to Optional[UploadFile] = None (no File() wrapper) so the
field is genuinely optional for text-input turns.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Frontend:
- Add pill-shaped text input + send button (→) to the left of the mic button
- Enter key or → click sends text directly without recording audio
- Input is disabled while a turn is processing; cleared on submit
- Welcome message updated to mention both input methods
- New CSS: .conv-input-bar, .conv-text-row, .conv-text-inp, .conv-send-btn,
.conv-divider (visual separator between text and mic sections)
Backend:
- /api/conversation/turn: audio is now optional (UploadFile | None)
- New text form field — when provided, STT step is skipped and text is
used as the transcript directly; SSE emits transcript event with stt_ms=null
- Raises 400 if neither audio nor text is supplied
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
