Thanks — this is a great question. You’re describing three related symptoms: speech that sounds too fast, parts of sentences being clipped/skipped, and odd intonation. I’ll explain what’s likely happening and give clear, actionable steps you can try. What’s likely happening • Streaming can cause rushed or clipped speech when TTS starts speaking before the LLM has emitted a full, natural text chunk. If the TTS receives very small fragments or tokens, it may begin speaking mid-phrase and miss later words or end sentences abruptly. • TTS settings (speaking rate / prosody controls) also directly affect speed and intonation. If rate is too high or no prosody/SSML hints are provided, speech will sound rushed and flat. • Some providers handle streaming and prosody better than others; provider-specific differences can make the same input sound very different. What you can try (short checklist) 1. Control chunking before TTS ◦ Ensure the LLM pushes complete sentences or logical phrases to the TTS (not tiny token fragments). ◦ Use sentence boundaries or punctuation to delay sending until a natural pause (periods, commas, newlines). ◦ If you have a custom

llm_node

, confirm it isn’t forwarding overly small chunks to

tts_node

. 1. Slow down the speaking rate ◦ Lower the TTS

speed

/

speaking_rate

(for example, try values around 0.8–1.0 as a starting point). ◦ Different plugins name this differently; set the provider-specific rate parameter. 1. Improve prosody / intonation ◦ Use SSML or provider voice instructions (emotion, emphasis, pauses) to guide tone and rhythm. ◦ Send structured output from the LLM that includes voice instructions or break markers the TTS can consume. 1. Try provider differences to isolate cause ◦ Test the same input against another TTS provider (e.g., ElevenLabs, Cartesia, etc.). If another provider sounds correct, the issue is likely provider-specific configuration rather than streaming logic. ◦ Some providers support word-level timing or richer SSML which helps when streaming. 1. Enable better alignment if available ◦ If your setup supports options like

use_tts_aligned_transcript

(or equivalent), enable it to improve sync and reduce clipping. How to proceed (quick troubleshooting flow) 1. Send a complete sentence chunk (with punctuation) and a slower

speaking_rate

. See if clipping stops. 2. If clipping continues, switch providers for a quick A/B test. If the other provider behaves, tune the original provider’s prosody/SSML options. 3. If speech is still unnatural, include explicit SSML breaks/pauses in the text the TTS receives. If you want, tell me: • Which TTS provider(s) you’re using now (plugin name), • How you’re streaming (are you sending sentence-delimited chunks or token-by-token?), and • The typical `speed`/`speaking_rate` value you’re using. With that, I can suggest exact parameter names and values or sample SSML snippets to try. Sources: Live Conversations with AI using ChatGPT and WebRTC | Text-to-speech (TTS) models overview | LiveKit Documentation | Tool definition and use | LiveKit Documentation | Pipeline nodes and hooks | LiveKit Documentation | MiniMax TTS plugin guide | LiveKit Documentation