Default sentence tokenizer for a provider's streamed TTS input.
The provider's hard max chunk length caps every emitted token. When expressive is set,
it also raises the minimum to EXPRESSIVE_BATCH_LEN so a couple of consecutive
sentences ride one request, keeping prosody continuous across the turn; otherwise tokens
emit per sentence (the unchanged default). Providers with no configured limit are
uncapped and stay per-sentence even under expressive — Fish Audio's markers are
sentence-scoped, so batching would cost time-to-first-audio and buy no steering.
Default sentence tokenizer for a provider's streamed TTS input.
The provider's hard max chunk length caps every emitted token. When
expressiveis set, it also raises the minimum to EXPRESSIVE_BATCH_LEN so a couple of consecutive sentences ride one request, keeping prosody continuous across the turn; otherwise tokens emit per sentence (the unchanged default). Providers with no configured limit are uncapped and stay per-sentence even under expressive — Fish Audio's markers are sentence-scoped, so batching would cost time-to-first-audio and buy no steering.