Feature Request: Add Voice Similarity Parameters Support to ElevenLabs Plugin

[chauhansarthakiitd]

Request support for ElevenLabs voice consistency parameters (seed, previous_request_ids, next_request_ids) in LiveKit's ElevenLabs TTS plugin for improved voice continuity.
Missing Parameters

seed - Deterministic voice generation (integer 0-4294967295)
previous_request_ids - Voice continuity from previous requests (max 3 IDs)
next_request_ids - Voice continuity for upcoming requests (max 3 IDs)
previous_text/next_text - Text context for prosody flow

Current Implementation Analysis
Looking at the source code in tts.py, the _TTSOptions dataclass and constructor currently support:

✅ voice_id, voice_settings, model, encoding
✅ streaming_latency, enable_ssml_parsing, language
❌ Missing voice similarity parameters listed above

Use Case
Essential for maintaining voice consistency in:

Long-form content split across multiple requests
Conversational AI with consistent voice across turns
Content regeneration with similar characteristics

Implementation
Add missing parameters to _TTSOptions dataclass and TTS constructor, then include in WebSocket/API calls to ElevenLabs.

Priority
Medium-High - This would help maintain voice consistency in conversational AI applications where users expect natural, continuous speech across multiple interactions. ElevenLabs sometimes fluctuates significantly when these parameters are not passed, to the point where alternate turns could have different voices talking (although not for all voice IDs).

References
LiveKit ElevenLabs TTS Source Code
ElevenLabs TTS API
Best practices for large text

[ashraq1455]

+1 — It would be great to have these parameters added soon. In my case, the buffering on ElevenLabs’ side when using ws (mentioned here in docs) introduced too much latency, so I wrote a custom tts_node that uses the ElevenLabs plugin’s synthesize method over the HTTP streaming endpoint (see docs). Since then, I’ve noticed the audio quality can vary quite a bit between chunks/sentences, sometimes significantly. These parameters would help reduce that inconsistency.

[davidzhao]

This isn't necessary since our plugin is using the streaming endpoint. When streaming, consistency is automatically handled by 11labs since it's a single connection.

within the framework, we do not use the http endpoint to synthesize as it's not recommended by them. LLMs generate tokens progressively and it'd be better to send it all to the TTS provider as they come in.