"""Speaker-side streaming pipeline for ``tools.tts_tool.stream_tts_to_speaker``. Turns a queue of LLM text deltas into audio the moment each sentence is complete. Two paths share the sentence cutter (``tools.tts_streaming``): :class:`_StreamerPlayback` for a registered chunked streamer (prefetch thread per sentence, one FIFO playback worker through a sounddevice OutputStream or temp WAV + system player) and :class:`_SyncSentencePipeline` for every other provider (per-sentence ``text_to_speech_tool`` on a single-thread executor, overlapped with playback). Seams tests monkeypatch on the origin module are resolved through :func:`_origin` at call time. """ from __future__ import annotations import logging import os import platform import queue import tempfile import threading from concurrent.futures import Future, ThreadPoolExecutor from typing import Callable, Iterable, Iterator, List, Optional logger = logging.getLogger("tools.tts_tool") def _origin(): from tools import tts_tool return tts_tool def _unlink_quietly(path: Optional[str]) -> None: if path: try: os.unlink(path) except OSError: pass def _align_int16_chunks(chunks: Iterable[bytes], stop_evt: threading.Event, *, pad_tail: bool = True) -> Iterator[bytes]: """Yield int16-aligned byte chunks; a dangling odd byte is padded at the end (or dropped).""" leftover = b"" for chunk in chunks: if stop_evt.is_set(): break buf = leftover + chunk aligned_len = len(buf) - (len(buf) % 2) if aligned_len >= 2: yield buf[:aligned_len] leftover = buf[aligned_len:] if aligned_len < len(buf) else b"" if leftover and pad_tail: yield b"\x00" def _play_via_tempfile(audio_iter: Iterable[bytes], stop_evt: threading.Event, sample_rate: int = 24000) -> None: """Write PCM chunks to a temp WAV file and play it with the system player.""" tmp = None tmp_path = None try: import wave tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False) tmp_path = tmp.name with wave.open(tmp, "wb") as wf: wf.setnchannels(1) wf.setsampwidth(2) # 16-bit wf.setframerate(sample_rate) for aligned in _align_int16_chunks(audio_iter, stop_evt): wf.writeframes(aligned) # wave.open() on a file object does NOT close it; on Windows the open write # handle blocks the player and the unlink below (WinError 32). tmp.close() from tools.voice_mode import play_audio_file play_audio_file(tmp_path) except Exception as exc: logger.warning("Temp-file TTS fallback failed: %s", exc) finally: if tmp is not None: try: tmp.close() # idempotent; ensures close on early error except Exception: pass _unlink_quietly(tmp_path) def _drain_chunks(chunk_queue: "queue.Queue[Optional[bytes]]") -> List[bytes]: """Collect one sentence's PCM chunks up to the ``None`` sentinel.""" chunks: List[bytes] = [] while True: chunk = chunk_queue.get() if chunk is None: return chunks chunks.append(chunk) class _SyncSentencePipeline: """Overlap per-sentence synthesis with playback for non-streaming providers. One single-thread synthesis executor (FIFO; providers never see concurrent calls) feeds one playback worker through a small bounded queue, so sentence n+1 synthesizes while n plays; the bound keeps lookahead/temp files small and gives the caller backpressure. ``text_to_speech_tool`` / ``play_audio_file`` are resolved late so test patches apply.""" def __init__(self, stop_event: threading.Event, *, lookahead: int = 2): self._stop = stop_event self._queue: "queue.Queue[Optional[tuple[str, Future]]]" = queue.Queue(maxsize=max(1, lookahead)) self._executor = ThreadPoolExecutor(max_workers=1, thread_name_prefix="tts-sync-synth") self._player = threading.Thread(target=self._drain, name="tts-sync-play", daemon=True) self._player.start() def speak(self, cleaned: str) -> None: """Queue one sentence. Blocks only when the lookahead bound is full.""" if self._stop.is_set(): return self._queue.put((cleaned, self._executor.submit(self._synthesize_to_tmp, cleaned))) def close(self) -> None: """Flush queued sentences in order (skipped if stopped), then join.""" self._queue.put(None) self._player.join() self._executor.shutdown(wait=True) def _synthesize_to_tmp(self, cleaned: str) -> Optional[str]: if self._stop.is_set(): return None tmp_path = None try: fd, tmp_path = tempfile.mkstemp(suffix=".mp3") os.close(fd) _origin().text_to_speech_tool(text=cleaned, output_path=tmp_path) return tmp_path except Exception as exc: logger.warning("Sync per-sentence TTS synthesis failed: %s", exc) _unlink_quietly(tmp_path) return None def _drain(self) -> None: while True: item = self._queue.get() if item is None: return _sentence, future = item tmp_path = None try: tmp_path = future.result() if tmp_path and not self._stop.is_set() and os.path.isfile(tmp_path) and os.path.getsize(tmp_path) > 0: from tools.voice_mode import play_audio_file play_audio_file(tmp_path) except Exception as exc: logger.warning("Sync per-sentence TTS failed: %s", exc) finally: _unlink_quietly(tmp_path) class _StreamerPlayback: """Prefetch + FIFO playback for a chunked :class:`StreamingTTSProvider`. ``speak(text)`` starts ``streamer.stream()`` immediately on a prefetch thread (at most 3 in flight) buffering into a bounded per-sentence queue; one playback worker drains those in order, so sentence N+1 arrives while N plays. Output is a PortAudio stream when one opened, else temp WAV files; a failing write is retried on a reinitialized stream up to ``_MAX_REINIT`` times before falling back to temp files.""" _MAX_REINIT = 3 _CHUNK_QUEUE_MAX = 64 def __init__(self, streamer, stop_event: threading.Event): self.streamer = streamer self.stop_event = stop_event self.output_stream = self._open_output_stream() self._audio_queue: "queue.Queue[Optional[queue.Queue[Optional[bytes]]]]" = queue.Queue() self._prefetch_threads: List[threading.Thread] = [] self._prefetch_sem = threading.Semaphore(3) self._worker = threading.Thread(target=self._playback_worker, daemon=True) self._worker.start() # -- PortAudio stream management --------------------------------------- def _create_output_stream(self): sd = _origin()._import_sounddevice() stream = sd.OutputStream(samplerate=self.streamer.sample_rate, channels=self.streamer.channels, dtype="int16") stream.start() return stream def _open_output_stream(self): # On macOS skip sounddevice entirely: PortAudio/CoreAudio init triggers a # kTCCServiceMediaLibrary permission prompt even though output needs no # media-library access. None routes every sentence through tempfile -> afplay. if platform.system() == "Darwin": return None try: return self._create_output_stream() except (ImportError, OSError) as exc: logger.debug("sounddevice not available, streamer→tempfile: %s", exc) except Exception as exc: logger.warning("sounddevice OutputStream failed: %s", exc) return None def _reinit_output_stream(self): """Close the broken PortAudio stream and try to create a fresh one.""" self.close_output_stream() try: self.output_stream = self._create_output_stream() logger.info("TTS: PortAudio output stream reinitialized after error") except Exception as exc: logger.warning("TTS: PortAudio stream reinit failed: %s", exc) self.output_stream = None return self.output_stream def close_output_stream(self) -> None: """Always release the device so a later stream can open it.""" if self.output_stream is not None: try: self.output_stream.stop() self.output_stream.close() except Exception: pass # -- prefetch ---------------------------------------------------------- def speak(self, text: str) -> None: """Start ``streamer.stream(text)`` and prefetch its chunks immediately.""" try: audio_iter = self.streamer.stream(text) except Exception as exc: logger.warning("Streaming TTS synthesis failed: %s", exc) return self._prefetch_sem.acquire() chunk_queue: "queue.Queue[Optional[bytes]]" = queue.Queue(maxsize=self._CHUNK_QUEUE_MAX) self._audio_queue.put(chunk_queue) t = threading.Thread(target=self._consume_to_queue, args=(audio_iter, chunk_queue), daemon=True) self._prefetch_threads.append(t) t.start() def _consume_to_queue(self, audio_iter: Iterator[bytes], chunk_queue: "queue.Queue[Optional[bytes]]") -> None: try: for chunk in audio_iter: if self.stop_event.is_set(): logger.info("TTS CUT: prefetch cancelled (stop_event set mid-sentence) — partial audio only") break chunk_queue.put(chunk, timeout=30.0) except Exception as exc: logger.warning("TTS CUT: streaming TTS prefetch failed mid-sentence (partial audio only): %s", exc) finally: chunk_queue.put(None) # sentinel: no more chunks self._prefetch_sem.release() # -- playback ---------------------------------------------------------- def _play_sentence_via_tempfile(self, chunk_queue) -> None: _play_via_tempfile(iter(_drain_chunks(chunk_queue)), self.stop_event, self.streamer.sample_rate) def _for_each_sentence(self, play: Callable[[queue.Queue], None]) -> None: """Feed queued sentences to *play* in order until the end sentinel; stopped sentences are skipped.""" while True: chunk_queue = self._audio_queue.get() if chunk_queue is None: return if not self.stop_event.is_set(): play(chunk_queue) def _write_pcm(self, buf: bytes) -> None: self._current_stream.write(self._np.frombuffer(buf, dtype=" bool: """Reinit the PortAudio stream after a failed write; False once ``_MAX_REINIT`` is exhausted.""" if self._reinit_count < self._MAX_REINIT: self._reinit_count += 1 self._current_stream = self._reinit_output_stream() return self._current_stream is not None logger.warning( "TTS: PortAudio reinit exhausted after %d attempts, falling back to tempfile for remaining sentences", self._MAX_REINIT, ) self._current_stream = None return False def _play_sentence_via_stream(self, chunk_queue) -> None: """Write one sentence's PCM to PortAudio; after an unrecoverable write failure the rest is dropped.""" if self._current_stream is None: self._play_sentence_via_tempfile(chunk_queue) return for aligned in _align_int16_chunks(iter(chunk_queue.get, None), self.stop_event, pad_tail=False): try: self._write_pcm(aligned) except Exception as write_exc: logger.warning("PortAudio write failed, attempting stream reinit: %s", write_exc) if not self._recover_stream(): return try: self._write_pcm(aligned) except Exception: pass def _playback_worker(self) -> None: """Single consumer: play audio segments from the queue in order.""" if self.output_stream is None: self._for_each_sentence(self._play_sentence_via_tempfile) return import numpy as _np try: from tools.voice_mode import mark_audio_output_active except Exception: def mark_audio_output_active(_active): return None self._np = _np self._reinit_count = 0 self._current_stream = self.output_stream mark_audio_output_active(True) try: self._for_each_sentence(self._play_sentence_via_stream) finally: mark_audio_output_active(False) def finish(self) -> None: """Send the end sentinel, then wait for playback and prefetch threads.""" self._audio_queue.put(None) self._worker.join(timeout=300.0) for t in self._prefetch_threads: t.join(timeout=10.0) self.close_output_stream() def stream_tts_to_speaker( text_queue: queue.Queue, stop_event: threading.Event, tts_done_event: threading.Event, display_callback: Optional[Callable[[str], None]] = None, provider: Optional[str] = None, ): """Consume text deltas from *text_queue*, cut into sentences, speak each the moment it's ready. A registered streaming provider plays chunked PCM; every other provider is spoken per-sentence via ``text_to_speech_tool``, so audio still starts on sentence one. Protocol: ``str`` deltas, a ``None`` sentinel = end-of-text (flush), *stop_event* aborts (barge-in), *tts_done_event* is **set** in ``finally`` so continuous voice mode knows playback finished.""" tts_done_event.clear() origin = _origin() sync_pipeline: Optional[_SyncSentencePipeline] = None playback: Optional[_StreamerPlayback] = None try: tts_config = origin._load_tts_config() # Prefer a chunked streamer for low time-to-first-audio; otherwise per-sentence # sync synthesis (universal — edge + every non-streamer). from tools.tts_streaming import SentenceChunker, resolve_streaming_provider streamer = resolve_streaming_provider(tts_config, preferred=provider) stream_max_len = 0 if streamer is None: sync_pipeline = _SyncSentencePipeline(stop_event) else: try: stream_max_len = origin._resolve_max_text_length(provider or origin._get_provider(tts_config), tts_config) except Exception: stream_max_len = 0 playback = _StreamerPlayback(streamer, stop_event) chunker = SentenceChunker() long_flush_len = 100 queue_timeout = 0.5 spoken_sentences: list[str] = [] # skip duplicate/near-duplicate sentences (LLM repetition) def _speak_sentence(sentence: str) -> None: if stop_event.is_set(): return cleaned = origin._strip_markdown_for_tts(sentence).strip() if not cleaned: return cleaned_lower = cleaned.lower().rstrip(".!,") if any(prev.lower().rstrip(".!,") == cleaned_lower for prev in spoken_sentences): return spoken_sentences.append(cleaned) if display_callback is not None: display_callback(sentence) # raw sentence on screen before TTS processing if sync_pipeline is not None: sync_pipeline.speak(cleaned) return if stream_max_len and len(cleaned) > stream_max_len: cleaned = cleaned[:stream_max_len] playback.speak(cleaned) while not stop_event.is_set(): try: delta = text_queue.get(timeout=queue_timeout) except queue.Empty: # Idle producer: flush a long buffer instead of sitting on it if len(chunker.buf) > long_flush_len: for sentence in chunker.flush(): _speak_sentence(sentence) continue if delta is None: for sentence in chunker.flush(): _speak_sentence(sentence) break for sentence in chunker.feed(delta): _speak_sentence(sentence) while True: try: text_queue.get_nowait() except queue.Empty: break except Exception as exc: logger.warning("Streaming TTS pipeline error: %s", exc) finally: # Flush the sync pipeline first: queued sentences finish playing (or are skipped # when stop_event is set) BEFORE tts_done_event fires, so continuous voice mode # never reopens the mic over its own voice. if sync_pipeline is not None: try: sync_pipeline.close() except Exception: pass # The end sentinel lives in finally: so an exception in the text pump still lets # the playback worker exit. if playback is not None: playback.finish() tts_done_event.set()