Reachy Conv

Fix: spontaneous voice never fired + idle random moves only fired when disarmed. Two bugs from v0.5.54/v0.5.55. (1) The spontaneous-voice loop's idle guard checked self._buffer to mean "the user is talking" — but every mic frame is appended to that buffer while armed (silence included), so it was non-empty the entire armed-silent time and the guard gated the loop off forever. The robot never spoke on its own, no matter how long you waited. Fix: guard on self._vad_has_speech (True only between VAD speech-start and speech-end — the real "you're talking right now" signal). (2) The idle random moves (fidget) were gated to MotionState.IDLE only, which is emitted only on disarm — so they never fired while the robot was armed and silent on the desk (the way you actually use it). Fix: also allow MotionState.LISTENING (armed + silent) in the dispatcher gate. Fidget now fires whenever the robot isn't in a turn — armed-silent or disarmed — and stays suppressed during think / speak / speak-done and while you're actually talking (VAD). The interval clock keeps accumulating across IDLE ↔ LISTENING so it fires as long as the app is up.

Motion tab — spontaneous voice interjection (the "alive on the desk" feature). New "Spontaneous voice" card: the voice counterpart of the idle random moves. Enter up to 12 text prompts (in a collapsible list — click to open/close them all), set an interval (30–3600 s, default 180 = once per 3 min), and enable. While the robot is armed and truly idle, it picks one prompt at random, feeds it to the LLM as if you spoke it, and speaks the reply out loud (asks you something, tells a fact, …). It then re-arms the mic so you can answer — but the re-arm respects the orb mute: muted → armed but deaf (you can't talk back); unmuted → you can reply. It never fires when disarmed (standby), while you're talking, or while it's speaking — and it can't overlap itself or a real turn (guarded by the existing listening/speaking signals). The injected line shows in the chat and the LOGS tab. Default off = zero behavior change. All live-apply + persist; restart-safe.

Motion tab — idle random moves + breathing auto-restart toggle. New "Idle random moves" card: pick up to 8 gestures from the emotion library, set an interval (10–1800 s, default 120 = once per 2 min), and enable. While the robot is truly idle (procedural breathing in IDLE — never during listening / thinking / speaking), it pauses the breathing, plays one randomly-picked move, then resumes breathing immediately (no 30 s dead gap). The interval clock resets on every state change so a conversation doesn't eat into it. Default off = zero behavior change. Also a new "Idle breathing auto-restart" checkbox: off = breathing keeps running but stops re-creating itself on every state transition (the "why does it restart when it's already breathing" blip is gone); on = re-starts as before. Breathing is never stopped by this toggle — it only gates the re-start-when-already-working. All live-apply + persist; restart-safe.

Fix: barge-in (voice interrupt) left the robot dead — stuck in listening forever. Interrupting the robot mid-speech primed the barge-in VAD but never primed the listening VAD with the lead-in speech, so it had no record of speech and could never fire utterance_ended → the buffer never flushed → the robot stopped responding. Fix: after interrupt(), prime the listening VAD with the barge-in lead-in frames so it can endpoint normally.

The loaded quant is now logged. The qwentts model: loaded ... line includes quant=Q4_K_M|Q8_0|BF16|F32 so you can see which quantization is actually running (matches the TTS-tab dropdown).

Qwen TTS settings panel — quant dropdown + sampling knobs + in-tab preview. A new "Qwen TTS settings" card in the TTS tab (shown only when engine = qwentts_cpp) exposes the real audio-quality levers, styled like the faster-qwen3-tts demo but native HTML/JS — no Gradio, no second process, one engine. Quantization dropdown (Q4_K_M / Q8_0 / BF16 / F32) picks the GGUF level — higher = cleaner vocoder reconstruction, more VRAM. Switching unloads the old model first (frees VRAM before the new one loads, via the existing _do_load chokepoint), then loads the new quant. Default stays Q4_K_M (zero behavior change on existing installs — you opt into a higher quant by ear). The 7 sampling knobs (seed / max-new-tokens / top-k / temperature / top-p / repetition-penalty / do-sample) moved into the card; a one-click "Good defaults" button applies cooled sampling (temp 0.8 / top_p 0.9 / top_k 20 / rep 1.1) that kills the intermittent garble/noise. A preview player synthesizes sample text through the live engine (honors the quant + knobs just set) so you tune by ear without talking to the robot. All live-apply + persist; restart-safe.

Volume boost without the "crazy" voice — brickwall limiter replaces tanh. The old master-gain stage did x = np.tanh(x * gain), which engaged across the whole signal (even small values got compressed), so a 20 dB boost squashed everything into the cap — distorted/"crazy" voice and a loudness ceiling (20 dB wasn't actually louder than ~8 dB). Replaced with a soft-knee brickwall limiter: linear below the knee (full intonation, full gain — nothing touched), only peaks that would otherwise clip bend toward the ceiling. Net: louder + clean. No dynamics/compressor squashing (the 5 W speaker + maxed desktop-app volume now gets a real boost instead of being slammed into a tanh cap).

Fix: app crashed (exit 250 / CUDA out-of-memory) on long TTS sentences. The qwentts non-streaming synth path fell back to the binding's codec_chunk_sec=24 default, so the DAC vocoder decoded an entire long sentence as one giant chunk. A ~14 s / 166-frame sentence allocated a ~1.2 GB CUDA compute buffer, committed via ggml's VMM cuMemCreate — a sub-second spike past an 11 GB card shared with the LLM, even with ~2 GB of steady-state VRAM free (nvidia-smi's ~1 s poll can't see it). The spike aborted the process: CUDA error: out of memory → Fatal Python error: Aborted → exit 250. Pattern: one short reply speaks fine, then crash on the next reply containing a long run-on sentence (Hermes reasoning output is what exposed it). Fix: bound the vocoder chunk on the non-streaming path too, via the existing qwentts_codec_chunk_sec knob (default 1.0 s) — a 14 s sentence now decodes as 14 small ~86 MB chunks instead of one ~1.2 GB chunk. Audio is identical (chunks concatenate internally with 2 s left-context overlap, same as the streaming path that already did this and never crashed). The crash is the vocoder chunk size, not the gateway — any LLM (LM Studio / Hermes) that emits a long single sentence hit it; this fixes all of them. Streaming users were already safe; this makes batched playback safe too.

LLM idle-stream timeout default raised 8 s → 20 s (reasoning models). The watchdog force-closes the LLM stream when no token arrives for idle_stream_timeout_s. The old 8 s cap was too tight for reasoning models with reasoning baked in (can't be disabled — e.g. Qwen3.5-4B in LM Studio, run at "lower reasoning" via Hermes), which routinely have ~10 s+ gaps between streamed tokens while reasoning. 8 s cut the reply mid-sentence (the "...Each" cutoff: openai stream idle for 10.5s (cap 8.0s) — force-closing). 20 s is ~2× headroom over the observed gap and still well under the 300 s HTTP timeout; non-reasoning models are unaffected (gaps < 1 s). Bump to 30 in the dashboard if a harder reasoning step ever exceeds it.

Output voice EQ — in-app 10-band graphic EQ + end filters. Replaces the EasyEffects (Linux PipeWire) chain the robot used to shape its voice: the app now pushes PCM straight to the robot speaker, so a pure-numpy/scipy RBJ biquad chain does the same shaping in-app — identical on Linux / Windows / macOS, applied per chunk before push_audio_sample. 10 graphic peak bands (31 Hz, 62, 125, 250, 500, 1k, 2k, 4k, 6k, 8 kHz) as classic vertical faders — every fader works now (the old first band was a disabled high-pass; a dead slider). Separate end filters — high-pass, low shelf, high shelf — pulled out of the graphic row into their own labeled panel (still user-tunable, just not disguised as a band). Master gain + tanh soft-clip so a boost doesn't hard-clip. Presets (Clear & loud / Warm / Radio) move the 10 gains + the two shelf gains. Filter state carries across chunks (no edge clicks) and resets per reply.

The 16 kHz ceiling is the hardware, not a knob. Confirmed against the Reachy Mini SDK: the ReSpeaker XVF3800 DAC, the GStreamer capsfilter(rate=16000), and the .asoundrc dmix slave all pin the speaker to 16 kHz (AudioBase.SAMPLE_RATE = 16000). Nyquist = 8 kHz, so the top graphic band is 8 kHz — this is the full-range EQ for the robot; a 16 kHz band would be inaudible (the speaker can't reproduce it). The 48 kHz you may see in the SDK is the AEC (mic echo-cancellation) path, not the speaker.

LOGS tab now shows the full LLM reply. The conversation loop used to log a 200-char preview of the assistant answer ending in "…", so you couldn't see what the model actually said when debugging. It now logs the whole reply; the thinking length is still reported as a count alongside.

Every hard-coded tunable is now in the dashboard. New AUDIO tab (XVF3800 chip config — AGC / NS / NN / gamma / MGSCALE — with a live "Re-apply now" button, no restart). Advanced <details> cards in Conversation (mic gating / self-barge-in: post-TTS silence, barge lead-in, RMS threshold, PTT threshold), Motion (breathing z-amplitude / frequency / antenna sway, blend, idle delay, head-body yaw clamp), STT (partial-transcription window), and Hermes (stop timeout, log poll interval, log buffer size). Every new knob defaults to its old hard-coded value — zero behavior change unless you move it. config.yaml is left untouched; missing fields fall back to pydantic defaults.

LLM completion-api as a first-class engine. Separate CompletionApiClient (OpenAI-compat read surface: /v1/models list + ping) — not mixed into OllamaClient. The LLM-tab model picker and the STATUS probe now dispatch by llm.engine, so completion_api (LM Studio / llama.cpp / vLLM / OpenAI) lists its own models instead of silently showing a stale "saved, not on server" entry. The golden Ollama chat path is untouched.

Thinking on/off works on the completion-api path. think=false now sends reasoning_effort + (for the default none) chat_template_kwargs: {enable_thinking: false} — the levers LM Studio and vLLM actually honor (verified: qwen3.5 stops reasoning, gemma unaffected, no 400). New llm.reasoning_effort knob (none / minimal / low / medium / high) lets you match whatever the server accepts — switch to minimal for OpenAI direct so chat_template_kwargs is omitted and the 400 is avoided.

Mic mute / orb / daemon wobble. Orb click mutes/unmutes the mic (s2s onMicTap); arm/disarm has its own buttons. Head wobble uses the daemon-native robot.enable_wobbling (playback-synced) driven by the _speaking_until window, not the LISTENING state.

Natural human speech across all TTS engines: the full LLM reply is synthesized and played as one continuous utterance (batched playback) instead of per-sentence chunks, so prosody carries across sentence boundaries. Made the Reachy 16 kHz constraint explicit and added a warning on oversized rates. Hotfix line (v0.5.43.2 → v0.5.43.6): removed garbage files that broke the Windows install, and fixed a dashboard-save ReferenceError that broke every TTS engine except edge.

Motion tab play endpoint switched to path-parameter shape (POST /api/motion/play_emotion/{move_name}?loop=…) to match the official Reachy Mini Desktop App's pollen-robotics/reachy-mini-desktop-app Expression picker pattern (the daemon's own endpoint is POST /api/move/play/recorded-move-dataset/{ds}/{move}). No Pydantic body model — the v0.4.16 "Field required" 422 is now structurally impossible.

Cache-bust static assets (?v=<version> injected by index() into the <script>/<link> tags) so the browser is forced to fetch the new bundle on every reinstall. Diagnostic console.warn in playEmotionPreview logs the request body.

Dynamic /api/app/version from importlib.metadata.version(). The top-bar brand and About card now show the actually-installed version (was hardcoded "v0.1.0").

Defensive error-message coercion in api() and playEmotionPreview. Pydantic 422 detail lists are joined by ; instead of becoming "[object Object]".

Motion tab per-row Set button (focused PUT, hot-reloads the dispatcher's mapping on save, no restart) + real error feedback on Test (the r.ok === false toast that surfaced the silent "object = object" failure).

Auto-populate Ollama model dropdown on dashboard init; auto-save model on change. max_tokens raised to 600 for reasoning models.

Re-fetch Motion tab on every navigation (so the library list picks up the background download without a hard refresh); ↻ Retry library load button for failed downloads.

Emoji-strip toggle for TTS. Head-wobble fix on subsequent responses (was clipping when LLM replies were short).

LISTENING/IDLE = procedural BreathingMove; THINKING = one-shot. Smooth blend-in for every recorded move (so the head doesn't snap to the move's first frame). Motion tab dropdowns always editable even when the library isn't loaded.

State-driven emotion moves (GestureDispatcher + per-state map).

max_tokens raised to 200 (from 100) for reasoning models.

Bundle ffmpeg via imageio-ffmpeg (no system install required).

Wire edge-tts knobs (rate/volume/pitch) through settings, dashboard, and synth call.

Surface qwen→silero fallback reason; ship onnx-only silero VAD (dropped the silero-vad + torchaudio dep tree, which crashed on CPU-only robots without CUDA 13).