ElevenLabs v4 Turbo trial — Clara voice latency, 2026-09-28

1. TL;DR (updated 2026-09-29, after the Willows test)

• ElevenLabs shipped v4 Turbo on 2026-09-28. Same price as today's model. The speech engine starts faster (about 90 ms vs 157 ms).
• But callers do not feel it. On 30 real recordings at Willows the caller waits about 2.1 s for Clara's reply on v4 Turbo, against about 1.7 s on today's model in production. Voice quality is fine: no garbling, no dead air.
• Interruption is unproven. On Fede's own Willows call he talked over a 24-second answer and Clara neither stopped nor heard him. The robot cannot talk over Clara, so this needs a timed-audio test caller (a Twilio call from a tester number; needs Fede's go).
• Shipped and live: the speech model is now one global repo setting for all of PropFlow (never per customer), pinned to v3 everywhere except the Willows test fleet, which runs v4 Turbo. Switching everyone is a one-line change.
• Found along the way at Willows, unrelated to the speech model: an 11 to 17 s first leasing answer, a "let me pick up where we left off" filler at hand-off, a wrong first name in the greeting, and tour-booking tool failures. These need their own fixes.
• Also on this page: the 30-day guest card audit (section 5) and the expression-tags experiment (section 6).

2. What v4 Turbo is

• Released 2026-09-28. ElevenLabs claims about 150 ms median time to first speech and 90+ languages.
• Both eleven_v4 and eleven_v4_turbo are available on our account.
• Production Clara runs eleven_v3_conversational (voice Belle, stability 0.5, streaming latency 3, language model claude-sonnet-4-6).

Sources: TechCrunch launch coverage · ElevenLabs model docs

3. How we tested

4. Results

Judge pass / fail

The triage script is a known-behavior scenario that fails on today's production prompt too. It is not a model effect.

What the numbers mean

The voice engine is 55 ms faster per turn on v4 Turbo (139 to 83 ms mean). The language model takes about 950 ms on both, and it dominates what a caller hears, so the felt change is small. Audio since silence is higher in the v4 arm, which we did not expect.

Why audio-since-silence went up

We broke every v4 turn down into its parts: how long Clara waited after the caller stopped, the language model’s time to its first sentence, the voice engine’s time to its first byte, and whatever is left over.

• Not the wait after the caller stops. Median 0.29 s on v4, and only the first reply after the greeting is long.
• Not tool calls. The leasing calls made none; the triage calls made one hand-off each, 19–28 ms.
• Not the voice engine’s first byte. That fell to 83 ms, as advertised.
• The leftover. After subtracting all of the above, v4 Turbo still has a median 1.55 s (leasing 1.6 s, triage 1.0 s) before the caller hears audio. On v3 the same leftover was about 0.3 s. That is the whole regression.

The leftover comes in two sizes: about 0.4–1.0 s on turns that open with a full sentence, and about 1.5–1.9 s on turns that open with a short acknowledgement (“Great!”, “Perfect!”, “Got it”) followed by a question. Leasing is hit harder because its qualifying turns are mostly that shape. Working hypothesis, unproven: v4 Turbo reports its first byte on an early packet, then takes over a second to start real speech on a short fragment. Two ways to confirm: measure leading silence in the response audio, or merge the acknowledgement into the next sentence on one call. Transcripts showed no speech artifacts (no stage directions read aloud, no cut-off words); every call ended by the caller hanging up, none dropped.

5. Guest card audit, last 30 days (2026-08-29 to 2026-09-28)

A read-only count of AppFolio guest cards created in the window, for the companies that have data in production. Test and staff companies are excluded.

Takeaways

• 68% of cards are system-created from listing sites.
• Hand-typed cards come from “Assistant PM” (a shared login, 13), Kat Barker (11), Erika Wieser (3) and Clara herself (4).
• Situs and Yale 25 show zero because nothing mirrors their cards yet.

6. What else v4 brings, and the rollout plan

Facts below come from ElevenLabs' own API and docs unless marked otherwise. ElevenLabs publishes no migration guide from v3 to v4 and no end-of-life date for v3 Conversational; the only staged-migration advice found is third-party.

What is new in v4 / v4 Turbo

Expression tags experiment (arm C, 2026-09-29)

Fede asked us to try v4's inline direction tags in experimentation. Arm C = the same two clones on v4 Turbo with expressive mode on and four tags offered (reassuring, warm, apologetic, pause), six robocalls, same scripts as arm B.

Dead air did not come back. Every call's audio was re-transcribed and scanned for pauses of 1.2 s or more inside Clara's turns: none. No tag word was spoken aloud. That was the failure that retired expression on v3 in August, and v4 Turbo did not reproduce it.

But the model never used a tag. No agent line in any of the six calls carried a bracket cue, so there is no evidence yet that the tags change how Clara sounds. Six calls is also too few to read the judge difference as real. Next: a larger batch with a prompt nudge that invites the tags, and Fede's listen on the bound line. Until then expression stays off in the repo; the four-tag change is uncommitted in an experiment worktree only.

Runner changes shipped in PR 13529 (an --apply-sound flag and per-call transcript saving, which closes the gap where deleted clones lost their conversations).

Rollout plan, three steps

Fede's standing goal for every voice change: more natural, faster, and at least as precise as today.

7. Willows proof on the real test line (2026-09-29)

Verdict: hold production. On real recordings, callers wait about 0.4 s longer for Clara on v4 Turbo than on today's model. The speech engine itself is faster, but the caller does not feel it. Interruption could not be proven either way.

What ran: 30 graded robot calls into the Willows line (v4 Turbo), 128 Clara turns, pause measured from the call audio (end of the caller's last word to Clara's first word). Comparison: 24 real production calls on today's model (68 turns). The comparison is indicative, not like for like: production callers are people, Willows callers were the robot.

• The overnight "1.5 s gap" was mostly a measurement effect. The vendor timer overstates the real pause by about 0.8 s on v4 Turbo and 0.2 s on v3. The audible difference is about 0.4 s, and it is the same on short and long replies.
• Interruption: not proven. On Fede's own call (14:28Z) he spoke over Clara's 24-second first leasing answer; his words are audible in the recording but Clara neither stopped nor transcribed them. Settings match production except the speech model. The robot cannot talk over Clara, so the ten talk-over calls produced no overlap. A timed-audio test caller is needed to settle it.
• Judge: 2 of 30 passed, but the failures are about what Clara says and does, not how she sounds, and the triage script also fails on v3. Found along the way at Willows, independent of the speech model: the first leasing answer runs 11 to 17 s; a filler line "let me pick up where we left off" plays at hand-off (13 times here, never in production); the greeting used the wrong first name; the booking tool failed or was walked back in 9 calls.

Files: ~/agents/006/v4-turbo-trial-2026-09-28/willows-proof/ (report, 40 call recordings and transcripts, analysis scripts). Runner: PR 13746.

8. Decisions for Fede

The production switch, when it comes, is a one-line change to the global setting, shipped through the normal merge-to-main sync, and reverted the same way.

9. Artifacts

• Test runner: PR 13472
• Run ids: 2026-09-29T01-23-46-731Z (Arm A, v3) and 2026-09-29T01-37-44-289Z (Arm B, v4 Turbo)