Eleven v4 and v4 Turbo serve different speech workflows. Evaluate pronunciation and full interaction delay separately from expressive delivery or a vendor inference-latency figure.
Two named models, two application paths
ElevenLabs launched the family on September 28. Documentation lists eleven_v4 for expressive speech and eleven_v4_turbo for real-time use, both with more than 90 languages. The changelog identifies the Text to Dialogue API and a WebSocket path for interactive applications. Its roughly 100 ms Turbo figure is median inference latency, not complete microphone-to-playback delay. Provider preference tests are not independent results measured by this publication.
A voice trial needs difficult words
Our recommended audition script includes a name, a street address, an acronym, a number and a short emotional transition. Have a fluent reviewer listen for meaning and pronunciation, not merely a pleasant timbre. Keep the approved transcript beside the audio. For a customer-facing application, one misread date or amount can matter more than the model’s expressive range. Reject outputs that convey the wrong instruction even when they sound polished.
Measure the entire conversational turn
For a live assistant, measure from the user finishing speech to the first audible response. Separate capture, transcription, model reasoning, synthesis, buffering and playback in the timing record. Try interruption and a reconnect, not only the easiest uninterrupted exchange. These are our engineering acceptance criteria, not vendor specifications. If the assistant answers quickly but cannot stop when interrupted, its nominal synthesis speed is not enough to make the interaction useful.
Expressive controls need editorial restraint
Our recommendation for an informational voice is a neutral baseline with limited emphasis. Preview emotional tags on the actual words rather than assuming one style works everywhere. A reassuring tone should not make an uncertain factual answer sound authoritative. Use voices only with the rights and consent required for the intended use, and make synthetic provenance clear where listeners could otherwise mistake it for a real speaker. Preserve the final script and approved render together.
Budget listening and correction, not only synthesis
Compare the expense and time required for an accepted finished clip, including retakes and human listening. Check current plan and endpoint billing instead of borrowing a rate from another model. This article does not establish a local downloadable checkpoint or GPU minimum, and no audio was generated for it. For an input-side contrast, the Whistle transcription guide and Phonon-2 guide cover recognition rather than synthesized speech. For the audition, add a sentence whose meaning changes with emphasis. Have the reviewer note whether the intended meaning survives, then try the same script in a second voice before making a production choice.
Frequently asked questions
Is the Turbo latency figure total conversation latency?
No. Keep inference latency separate from capture, transcription, reasoning, buffering and playback.
What should a voice audition include?
Names, numbers, abbreviations and emotional transitions, reviewed against the approved script.
Is speech recognition the same as voice generation?
No. Recognition turns audio into text; synthesis creates speech from an input representation.
Sources & further reading
Primary sources checked Oct 11, 2026. Vendor statements are attributed; editorial advice is our own.
- 1Introducing Eleven v4 ↗ElevenLabs
- 2Model IDs and capabilities ↗ElevenLabs
- 3September 28 API changelog ↗ElevenLabs
Help us keep this useful. Check contact availability →
