Buffer sizes, minimum sentence length, and related TTS streaming behaviors in LiveKit are configurable via parameters on the sentence tokenizer and the specific TTS provider implementation you use: • Sentence tokenizer settings: Most TTS providers use a sentence tokenizer (like NLTK or a basic tokenizer), which supports configuration of

min_sentence_len

(minimum sentence length in characters, default ~20) and

stream_context_len

(context length, default ~10). • TTS Provider options: Many providers (e.g., ElevenLabs, Google, Cartesia, Resemble) specify default buffered word counts, chunk length schedules, or allow you to pass a custom tokenizer object with your chosen settings. For instance, with ElevenLabs, you can set

chunk_length_schedule

and

auto_mode

(reduces buffering for lower latency). • Audio buffer size: If you encounter frame capture or latency issues, audio source configurations let you set

queue_size_ms

(recommendation: 1000-2000 ms for stability). Larger buffers reduce risk of frame errors but increase latency. Configuration is typically done via constructor parameters. Example for a sentence tokenizer:

tokenizer = SentenceTokenizer(
    min_sentence_len=25,
    stream_context_len=15
)
tts = SomeTTS(tokenizer=tokenizer)

Or with ElevenLabs:

tts = ElevenLabsTTS(
    chunk_length_schedule=[80, 120, 200, 260],
    auto_mode=True
)

The right balance of buffer size and minimum sentence length depends on your latency/performance preferences. Let me know which provider you're using or if you need a code example for a specific environment. Sources: livekit-plugins/livekit-plugins-elevenlabs/livekit/plugins/elevenlabs/tts.p... | SentenceTokenizer | LiveKit Agents | ElevenLabs integration guide | LiveKit Docs | Speech & Audio Processing | livekit/livekit_composite | DeepWiki