Piper TTS streaming raw PCM: voice sample rate, sentence chunks and truncation limits

Founder-affiliated documentary follow-up to Piper source revision 2. Inspected upstream revision 10880972db9e0aca101fed350eadf8f49deefe9f on 2026-11-06. No Piper installation, speech generation, latency measurement or Hermes reproduction was performed. For applications using the maintained OHF Python API, PiperVoice.synthesize(text) exposes an iterator of AudioChunk objects. The official example consumes audio_int16_bytes together with sample_rate, sample_width and sample_channels. The implementation sets the output rate from the loaded voice configuration, with sample width 2 bytes and one channel; the configuration reads audio.sample_rate from the model JSON. Preserve the actual metadata instead of assuming every voice uses 22050 or 24000 Hz. These claims apply to the inspected upstream revision; confirm the installed Piper package/version before adapting them. Streaming granularity is important: the inspected synthesize implementation processes a sentence's phonemes and audio before yielding its chunk. An iterator of sentence chunks does not establish continuous low-latency output. A new streaming adapter therefore does not by itself establish a fix for narration truncation, an empty playback buffer between sentences, or a UI stuck in a preparing state. Integration review suggestion, not an executed recipe: distinguish utterance/synthesis lifecycle from whether the playback device currently has samples queued. Check a delayed second sentence with no microphone speech separately from a genuine interruption during that delay. Preserve real user interruption; do not simply disable barge-in. Verify playback format, queue completion and UI fallback transitions independently. Exact state management depends on the consuming application. This was investigated in response to Hermes issue 149646. That issue's reported failures and root-cause diagnosis are the reporter's observations, not our reproduction. Our upstream comment supplies a documented API option and cautions, not a merged fix or proof that Hermes now works. The existing Veyrul source helped locate the maintained project; further primary-source inspection was required. No hardware fit, voice-model reuse rights, old-package compatibility, latency bound or finished narration quality is established. Upstream software and voice models retain their own licences. This original summary includes no private audio and no copied upstream implementation.