Chunked synthesis splits long texts into sentences and pipelines
synthesis with playback — first audio plays within ~1-2s instead of
waiting for full synthesis. Queue supports suppress_exit_tone between
chunks for seamless playback.
Tool annotations (ToolAnnotations) signal behavioral hints to MCP
clients: speak() is openWorldHint=True (physical speakers),
cancel_speech() is destructiveHint=True, read-only tools marked
accordingly.
Progress notifications now include message= strings that tell LLM
clients they can continue working during playback. Server instructions
corrected from "returns immediately" to accurately describe blocking
behavior with parallel-safe guidance. speak() docstring updated to
encourage concurrent tool calls.
The same shutdown_timeout was used for both uvicorn handler drain AND
queue.stop() consumer drain — sequential phases sharing Docker's
wall-clock budget. With 30s each, worst case was 60s, exceeding
the 35s stop_grace_period and causing SIGKILL.
Fix: uvicorn gets a fixed 3s drain (handlers just re-raise), queue
gets the full shutdown_timeout. Docker grace = 3 + timeout + 5s safety.
Also adds shutdown observability: startup logs the timing chain,
queue.stop() logs remaining audio vs available budget.
speak()'s CancelledError handler was calling queue.cancel() which killed
pw-play mid-sentence during container restarts. Now it just re-raises,
letting the consumer finish naturally. Explicit cancel_speech() still
kills playback immediately.
Also override FastMCP's timeout_graceful_shutdown=0 in uvicorn config
so request handlers get shutdown_timeout seconds before cancellation.
speak() now blocks until playback finishes, reporting progress via SSE
(5% entry tone → 30% synthesis → 35-99% playing → 100% done). Entry
tone fires immediately on call to cover synthesis latency.
Queue shutdown waits for current speech to finish (configurable timeout,
default 30s) before draining pending items — no more mid-sentence
cutoffs on container restart.
Cancellation via cancel_speech() tool or MCP notifications/cancelled
kills pw-play and plays a vinyl scratch tone. Consumer continues to
next item after cancel.
Progress tracking uses a background ticker task instead of
asyncio.wait_for polling — the latter causes stale CancelledError
propagation to the consumer under Python 3.13.
Each project gets a distinct voice from a curated English pool,
assigned via round-robin with gender/accent interleaving for
maximum perceptual contrast between consecutive projects.
- New voice_identity.py: pool filtering, interleaving, persistence
- Round-robin replaces SHA-256 hashing (no collisions until pool
exhaustion at 22 voices)
- Assignments persist to /data/voice-assignments.json across restarts
- speak() and generate_audio() accept optional project= parameter
- MCP roots fallback with 2s timeout for future bidirectional clients
- English-only pool (af_/am_/bf_/bm_/ef_/em_ prefixes)
- af_nicole excluded from auto-assign (whispery), still explicit-ok
- Fix voice blacklist to use full identifiers (am_adam, af_jessica)
Queue-aware exit tones: descending roger beep when queue empties
("over"), ascending standby blip when more items are queued ("standby,
more coming"). Includes quindar-out (2475 Hz) as Apollo-themed
alternative. Configurable via TTS_EXIT_TONE env var.
Also adds CLAUDE.md documenting the full tone system.
Generate 48kHz WAV tones at startup (chirp and apollo/quindar) using
numpy. The queue consumer plays the selected tone before each speech
item via pw-play. Configurable via TTS_ENTRY_TONE env var: chirp
(default), apollo, none, or path to a custom WAV file.
SNAC's convolutional decoder has a receptive field spanning multiple
frames. Batched decode (28 tokens at a time) created boundary
discontinuities that produced muddled audio. Since pw-play waits for
the full WAV anyway, single-pass decode costs only ~2s extra and
produces clean audio.
Build llama.cpp from source with SM 120 CUDA kernels and FORCE_CUBLAS
for RTX 5070 Blackwell. Rewrite OrpheusEngine to stream tokens via SSE
and decode SNAC in overlapping 28-token batches (4 frames), replacing
the blocking requests+stream:false approach.
Performance: 13.5 → 170-213 tok/s. 100s audio generates in ~48s (2x
faster than realtime). Replaces requests with httpx async client.
Also switch MCP transport to stateless_http mode so container restarts
don't invalidate client sessions.
SNAC + torch no longer load at startup — deferred to first Orpheus
call via double-checked locking. Startup drops from ~13s to 0.5s,
idle RAM reduced by ~200MB. OrpheusEngine constructor no longer
takes snac_model; it self-loads on demand.
CPU-only torch image (~180MB vs 873MB CUDA), PipeWire socket
passthrough for audio playback, SNAC HuggingFace cache volume.
Served at voice.l.supported.systems via caddy-docker-proxy.
FastMCP 3.0 Streamable HTTP server exposing Piper (Wyoming/Docker),
Kokoro (ONNX), and Orpheus (Ollama+SNAC) as MCP tools. Includes a
FIFO speech queue so concurrent agents don't talk over each other —
waiting callers get queue position updates via ctx.info().
Tools: speak, generate_audio, list_voices, list_engines
Resource: audio://recent
