@@ -550,7 +550,7 @@ fn tts_worker(
550550 & mut first_append,
551551 audio,
552552 silence_buf_len,
553- model_chunk_index == 0 ,
553+ model_chunk_index == 0 || player . empty ( ) ,
554554 ends_playback_chunk,
555555 ) ;
556556
@@ -699,13 +699,12 @@ fn apply_fade_out(samples: &mut [f32]) {
699699 }
700700}
701701
702- /// Build the single buffer appended to the rodio `Player` for one synthesised
703- /// sentence.
702+ /// Build one buffer appended to the rodio `Player` for a synthesis unit.
704703///
705- /// Every sentence chunk gets a short lead-in pad immediately before its audio.
706- /// This matters for chunks that start with soft first phonemes (`I'm`, `I've`):
707- /// the synthesized buffer can begin with speech within the first millisecond,
708- /// so the playback layer must provide the device/mixer cushion.
704+ /// Every playback boundary gets a short lead-in pad immediately before its
705+ /// audio. This matters for chunks that start with soft first phonemes (`I'm`,
706+ /// `I've`): the synthesized buffer can begin with speech within the first
707+ /// millisecond, so the playback layer must provide the device/mixer cushion.
709708/// To keep the audible gap unchanged, the trailing silence after this chunk is
710709/// shortened by the same amount (`silence_buf_len - SENTENCE_LEAD_IN_SAMPLES`):
711710/// sentence N contributes 80 ms of post-speech silence and sentence N+1
@@ -716,15 +715,15 @@ fn apply_fade_out(samples: &mut [f32]) {
716715/// tracked source per synthesized sentence, avoiding source-boundary/drain
717716/// regressions from enqueueing the lead-in, audio, and tail as separate sounds.
718717///
718+ /// A playback chunk may contain several model-sized synthesis units. Only the
719+ /// first unit receives the onset cushion and only the last receives the
720+ /// remaining gap. If playback underruns while the next unit is synthesized,
721+ /// that unit becomes a new playback boundary and receives a fresh cushion.
722+ ///
719723/// `first_append` is flipped on the first call after the player goes idle.
720724/// The worker uses it in the idle branch of the main loop to distinguish
721725/// "never queued anything since last drain" from "drained after speaking",
722726/// which controls when `tts_active` is released and the lead-in re-armed.
723- /// Add silence only at the outer playback boundary.
724- ///
725- /// A playback chunk may contain several model-sized synthesis units. The first
726- /// unit receives the onset cushion, the last receives the remaining gap, and
727- /// intermediate units stay sample-contiguous.
728727fn build_sentence_append_buffer (
729728 first_append : & mut bool ,
730729 audio : Vec < f32 > ,
0 commit comments