#I'm trying to get consistent voice between requests, what are the optimal params to acheive this an

1 messages · Page 1 of 1 (latest)

rich prismBOT
#
AI-generated Answer for kes0953

Short answer

  • For consistency between requests in Eleven v3 TTS, fix the seed, keep voice settings constant (especially Stability high/“Robust”), avoid Style exaggeration, and use the same model/voice/text normalization. For multi‑chunk jobs, pass the continuity fields. Determinism is best‑effort only. (elevenlabs.io)

What to set (and why)

  • seed: Set a fixed integer seed so repeated requests with the same inputs attempt to return the same audio. Determinism isn’t guaranteed. Range: 0–4,294,967,295. (elevenlabs.io)
  • Stability (voice_settings.stability): Higher values yield more stable, consistent delivery; lower values add variation. In v3 specifically, choose the “Robust” stability preset for maximum consistency. Min/max not published in API docs. (elevenlabs.io)
  • Similarity (voice_settings.similarity_boost): Controls adherence to the original voice timbre. Keep constant across requests. Min/max not published in API docs. (elevenlabs.io)
  • Style exaggeration (voice_settings.style): Can make results less stable and adds latency; recommended to keep at 0 for predictable output. Min/max not published in API docs. (elevenlabs.io)
  • Speaker boost (voice_settings.use_speaker_boost): Subtly increases similarity; slightly higher latency. Keep constant if you use it. Boolean. (elevenlabs.io)
  • Speed (voice_settings.speed): Keep at default 1.0 for consistency; extreme values may affect quality. Range: 0.7–1.2. (help.elevenlabs.io)
  • Model choice: For maximum long‑form stability, Multilingual v2 is documented as “most stable” on long‑form; v3 is the most expressive. Pick one model and keep it fixed. (elevenlabs.io)
  • Text normalization: Leave apply_text_normalization on its default unless you need it; changing normalization can affect pronunciations between runs. Keep fixed across requests. (elevenlabs.io)

For multi‑request continuity (joining segments)

  • Provide previous_text/next_text or previous_request_ids/next_request_ids to help the model keep prosody consistent across clips. Use the same model for all segments. Max 3 request IDs each. (elevenlabs.io)

Documented parameter limits (v3 TTS API)

  • seed: integer 0–4,294,967,295. (elevenlabs.io)
  • speed: 0.7–1.2 (default 1.0). (help.elevenlabs.io)
  • stability, similarity_boost, style: documented as doubles, but the API reference doesn’t publish explicit min/max. Keep them constant; raise Stability and keep Style at 0 for predictability. (elevenlabs.io)

Notes

  • The models are nondeterministic by design; seed improves repeatability but subtle differences can still occur. (elevenlabs.io)

Relevant docs

  • Create speech (request body incl. seed, continuity fields, normalization): (elevenlabs.io)
  • Text to Speech overview (nondeterminism and seed): (elevenlabs.io)
  • Voice settings API (stability/similarity/style/speaker_boost/speed definitions):
gray pollen
rich prismBOT
#
AI-generated Answer for kes0953

Short answer: not for a single stitched chain. For stitching that uses previous_request_ids/next_request_ids, each referenced request must have fully finished processing before it can condition the next one—so you need to sequence those requests. For streaming, you must read the entire response body before using its request-id. (elevenlabs.io)

If you need parallelism, you can generate chunks concurrently using text conditioning (previous_text/next_text), since those don’t require completed prior generations. Best results typically come from conditioning on past generations (request IDs), or combining both approaches. (elevenlabs.io)

Other constraints to keep in mind:

  • Request IDs should be no older than about two hours. (elevenlabs.io)
  • Request stitching isn’t available for the eleven_v3 model. (elevenlabs.io)
  • In zero-retention mode (enable_logging=false; enterprise), history features like request stitching are unavailable. (elevenlabs.io)

Doc page you linked (How to use Request Stitching) for reference. (elevenlabs.io)

Sources:
Stitching multiple requests | ElevenLabs Documentation
Create speech | ElevenLabs Documentation
Stitching multiple requests | ElevenLabs Documentation
Stitching multiple requests | ElevenLabs Documentation
Create speech with timing | ElevenLabs Documentation
Stitching multiple requests | ElevenLabs Documentation

gray pollen
#

Will stitching be available in the future for v3? Why is it not available now?

rich prismBOT
#
AI-generated Answer for kes0953
  • Current status: Request Stitching is not available for the eleven_v3 model. This is explicitly called out in the Request Stitching guide. (elevenlabs.io)
  • Reason: The documentation does not provide a public explanation for why v3 doesn’t support stitching. (elevenlabs.io)
  • Future availability: There is no published ETA or commitment in the docs/blog for adding stitching to v3. You can monitor the v3 materials and the public changelog for updates. (elevenlabs.io)

Note: Stitching remains supported on other TTS models via previous_request_ids/previous_text/next_text, and it requires history features (logging) to be enabled. (elevenlabs.io)

Sources:
Stitching multiple requests | ElevenLabs Documentation
Stitching multiple requests | ElevenLabs Documentation
Eleven v3: Most Expressive AI Text to Speech Model Launched | ElevenLabs
Create speech | ElevenLabs Documentation

gray pollen
#

do 'voice_settings' have an effect when used with eleven_v3 model? Because in the eleven labs app there is only a stability setting?

rich prismBOT
#
AI-generated Answer for kes0953

Short answer: Only Stability is documented to affect the eleven_v3 (alpha) model. The v3 guide lists Stability as the key (and only) adjustable setting and directs you to control emotion/style and even speed via audio tags instead. The app showing only a Stability slider matches this guidance. (elevenlabs.io)

While the Text-to-Speech API still accepts a generic voice_settings object (stability, similarity_boost, style, use_speaker_boost, speed), ElevenLabs does not document similarity/style/speaker-boost as having any effect for v3. In other words, those fields are API-wide options, but only Stability is described for v3. Use audio tags for expressiveness and pacing. (elevenlabs.io)

Relevant docs:

  • Prompting Eleven v3 (alpha): Stability and audio-tag control (including speed). (elevenlabs.io)
  • TTS “Create speech” endpoint: voice_settings is accepted generically. (elevenlabs.io)
  • Voice settings API: lists available fields (model-agnostic). (elevenlabs.io)

If you need controls beyond Stability with v3, use audio tags (e.g., [whispers], [sarcastic], [excited]) in your text. (elevenlabs.io)

Sources:
Prompting Eleven v3 (alpha) | ElevenLabs Documentation
Create speech | ElevenLabs Documentation
Prompting Eleven v3 (alpha) | ElevenLabs Documentation
Create speech | ElevenLabs Documentation
Edit voice settings | ElevenLabs Documentation
Prompting Eleven v3 (alpha) | ElevenLabs Documentation

gray pollen
#

What is the max stability setting? 1?

rich prismBOT
gray pollen
#

For v3 with stability at 1.0, what is the optimal character length for my requests?

rich prismBOT
#
AI-generated Answer for kes0953

Short answer: Aim for roughly 250–800 characters per generation.

  • Eleven v3 guidance: very short prompts can be unstable; use prompts greater than 250 characters. (elevenlabs.io)
  • TTS best practice across models: keep each generation under 800–900 characters for best results; split longer text. (elevenlabs.io)
  • Hard limit: Eleven v3 supports up to 10,000 characters per request, but that’s a maximum, not an optimal target. (elevenlabs.io)

There’s no official guidance that stability = 1.0 changes these length recommendations; the above applies to v3 generally. (elevenlabs.io)

Sources:
Prompting Eleven v3 (alpha) | ElevenLabs Documentation
Troubleshooting | ElevenLabs Documentation
Text to Speech (product guide) | ElevenLabs Documentation
Prompting Eleven v3 (alpha) | ElevenLabs Documentation

gray pollen
#

I have generated 3000 character responses and they sound fine to me

rich prismBOT
#
AI-generated Answer for kes0953

That’s fine. For Eleven v3, a single text-to-speech request supports up to 10,000 characters, so 3,000 is well within the documented limit. (elevenlabs.io)

v3 also tends to perform better with non‑trivial prompts; Eleven recommends using prompts longer than ~250 characters, so your length is in the recommended range. (elevenlabs.io)

If you scale to longer passages, Eleven suggests splitting long text into segments and/or using streaming; Studio is recommended for very long‑form content. (elevenlabs.io, help.elevenlabs.io)

Note that at the “Robust” end (high stability), v3 trades some responsiveness to directional prompts for consistency; lowering stability increases expressiveness. There’s no published “optimal” character count beyond the limits above—use what sounds best for your use case. (elevenlabs.io)

Sources:
Text to Speech | ElevenLabs Documentation
Prompting Eleven v3 (alpha) | ElevenLabs Documentation
Text to Dialogue | ElevenLabs Documentation
What's the maximum amount of characters and text I can generate? – ElevenLabs
Prompting Eleven v3 (alpha) | ElevenLabs Documentation

gray pollen
#

I'm getting 429 errors, whats the limit on the grant/business plan?

rich prismBOT
#
AI-generated Answer for kes0953

Short answer: 429 means you’re hitting a concurrency cap (or the system is briefly busy). On Business, the documented concurrency limits are:

Notes:

  • 429 “too_many_concurrent_requests” = you exceeded your plan’s concurrency; 429 “system_busy” = transient load, retry with backoff. (help.elevenlabs.io, elevenlabs.io)
  • The public docs don’t list a separate “Grant” tier; limits are determined by the underlying plan on your account. You can confirm your exact cap via the response headers current-concurrent-requests and maximum-concurrent-requests. (elevenlabs.io)

Sources:
How many requests can I make and can I increase it? – ElevenLabs
Models | ElevenLabs Documentation
How many Speech to Text requests can I make and can I increase it? – ElevenLabs
How many Conversational AI requests can I make and can I increase it? – ElevenLabs
Models | ElevenLabs Documentation
API - Error Code 429 – ElevenLabs
Error messages | ElevenLabs Documentation
Models | ElevenLabs Documentation

gray pollen
#

WHat about for eleven_v3?