The Text-to-Speech endpoints do support previous_text and next_text as optional request fields to improve continuity between chunks. You can pass them in the JSON body of Create speech (and the streaming-with-timestamps variant). (elevenlabs.io)
However, the separate “request stitching” mechanism that conditions on previous_request_ids/next_request_ids is not available for the eleven_v3 model. (elevenlabs.io)
Eleven v3 can be used via the standard Create speech/Stream speech endpoints by setting model_id to eleven_v3. (help.elevenlabs.io)
If you still need to keep tone consistent across generations with v3, ElevenLabs’ docs recommend:
Use a fixed seed for best‑effort repeatability across runs (deterministic sampling when seed and parameters are identical, not guaranteed). (elevenlabs.io)
Tune voice settings for stability; v3’s “Stability” setting is the primary control for adherence/consistency (Robust = most stable). Avoid very short prompts—v3 is more consistent with longer inputs; the guide suggests >250 characters. (elevenlabs.io)
Provide explicit direction via audio/emotion tags and punctuation to keep delivery predictable; explicit tags yield more predictable results than relying only on context. (elevenlabs.io)
Summary:
Try passing previous_text/next_text on the TTS requests you’re making; they are documented fields on the endpoints. If you were relying on request stitching via previous/next_request_ids, note that it’s not supported on eleven_v3. Combine a fixed seed, higher stability, longer contiguous segments, and explicit tags to maintain a similar tone across generations. (elevenlabs.io)
sinful imp
hidden aurora