v0.5.43.6: hotfix dashboard save ReferenceError broke every TTS engine except edge

v0.5.43.5: batched playback (full reply as one continuous utterance)

v0.5.43.3: make Reachy 16kHz constraint explicit + warn on oversized rates

v0.5.43.2: natural human speech across all TTS engines

v0.5.43.1: fix silent LLM-reply audio — use shape-aware _synth_result_to_pcm in _synth_and_push_sentence

v0.5.43: wire nested gi into sys.path — real fix for v0.5.42 Windows install bug

v0.5.42: real gstreamer, no fake shim — auto-install on first run

v0.5.41: cover every gstreamer call site the SDK's WebRTC backend init makes

v0.5.40: force media_backend=webrtc when gi shim is active

v0.5.39: expand the gi shim's _GstModule to cover every Gst type the SDK references

v0.5.38: bundle the `gi` shim INSIDE the wheel as a real sub-package

v0.5.37: thinking filler plays the full phrase + reverted prod audio path

v0.5.36: TTS thinking filler on every backend + gi shim as fallback only + reachy-mini 1.9.0

v0.5.35: fix gstreamer isinstance crash in mic_pump (v0.5.34 regression)

v0.5.34: pre-register `gi` shim in sys.modules at top of main.py

v0.5.33: `gi` shim now overwrites partial `gi/` from gstreamer-python

v0.5.32: make `gi` shim probe a 2-line check (catches gstreamer-python partial-install)

v0.5.31: ship `gi` shim for Windows + bind pydub to imageio-ffmpeg at module-load

v0.5.30: dashboard SHUTDOWN button + clean app-close path (Ollama left alone)

v0.5.29.3: hotfix — don't wipe a valid ca_venv on Windows (WinError 32)

v0.5.29.2: hotfix — strip PYTHONPATH/PYTHONHOME when re-execing into ca_venv

v0.5.29.1: hotfix — call per-app venv bootstrap from main.py too

v0.5.29: per-app venv on first launch (Windows chatterbox fix)

v0.5.28: install-time numpy downgrade for Windows chatterbox

v0.5.27: auto-fix the numba/numpy 2.5 ImportError on Windows

v0.5.25: card list of cloned voices + turbo knob disabling + chatterbox dtype fix

v0.5.24: chatterbox engine save 500 + cloned voices actually used end-to-end

v0.5.22: chatterbox voice-clone upload accepts any audio format

v0.5.21: fat wheel for Windows install deadlock fix + qwentts dormant

v0.5.20: TTS-tab dashboard overhaul + thinking-filler audio + chunker mid-sentence fix

v0.5.19.3: pre-roll the first TTS sub-chunk while the LLM is still emitting

v0.5.19.2: pipeline edge-tts sub-chunks so the next WebSocket RTT runs in parallel

v0.5.19.1: edge-tts splits input into ~15-word sub-sentence chunks

v0.5.19: edge-tts synth_stream yields one chunk per sentence (was per-frame)

v0.5.18: pace TTS playback consumer at real-time + anti-aliasing resample

v0.5.16: chatterbox voice upload (WAV-only) + edge-tts synth_stream deadlock fix

v0.5.12: card-ify TTS page + skip undecodable edge-tts chunks

v0.5.11: edge-tts synth_stream uses run_coroutine_threadsafe (no frame drops)

v0.5.10: size playback_queue by max_inflight (chatterbox 1-sec-audio fix)

v0.5.8: lift arbitrary upper caps on streaming chunker knobs (book mode)

v0.5.7: expose streaming chunker knobs in the dashboard

v0.5.6: drop qwentts topbar pip and STATUS page row

v0.5.5: add fixed-length chunk ceiling for run-on LLM output with no punctuation

v0.5.4: parallel streaming TTS pipeline (xiaozhi/ElevenLabs pattern)

v0.5.3: streaming LLM + sentence-chunked TTS so audio starts the moment the first sentence is complete.

v0.5.3: streaming LLM + sentence-chunked TTS so audio starts the moment the first sentence is complete.

v0.5.2: fix the chatterbox import (torchvision hiding) + lock to CPU.

v0.5.1: fix the fresh-install conflict between chatterbox-tts and reachy-mini.

v0.5.0: add Chatterbox TTS (Resemble AI, MIT) as a new engine

v0.4.21: remove old +cpu wheels (had a + in the filename which Windows rejects); v0.4.21 uses -cpu wheels + stdlib zipfile extract at runtime, no pip install