Better voice. Better calls?

What changed

Announcement: 28 Sept 2026

ElevenLabs introduced Eleven v4 and its Turbo variant on 28 September 2026. The company describes more expressive speech, improved voice consistency and a Turbo model intended for lower-latency agent use. Its published speech timing excludes network latency, so it should not be treated as the response time of a complete phone workflow.

ElevenLabs’ release and measurement notes. Source checked 22 Oct 2026.

Test the whole call

Start with the caller’s task: book a service, confirm a time or ask for a callback. Measure from the end of the caller’s speech to the start of a useful response. Include speech recognition, reasoning, the calendar lookup, speech generation and the telephone connection.

Use the same scenarios for every shortlisted provider. Test a quiet call and a noisy one, an interruption, a correction and a request the agent cannot handle. Keep notes about what happened, rather than selecting a voice only because its demonstration sounded natural.

Protect the booking record

Our proposed booking workflow reads available slots from the scheduling system, asks the caller to choose, and writes the booking before confirming it aloud. If the write fails, the agent explains that the booking is not confirmed and offers a callback or another supported route.

Use a booking identifier across the scheduling tool, reminders and CRM. Deduplicate repeated events so a retry cannot create another appointment or another reminder call. A cancellation should stop outstanding reminders. These behaviours belong in the integration even if the voice provider changes.

Test Malayalam locally

For a Malayalam-speaking audience, use native-speaker review of the exact voices and speech models available to your account. Include Malayalam mixed with English, local place names, service terms, dates, phone numbers and callers who change their answer midway.

Do not assume a general multilingual claim proves acceptable Malayalam quality. Separate recognition errors from pronunciation problems and from incorrect business logic. That distinction tells you whether to change the speech component, the prompt, the reference information or the connected system.

Plan the human handoff

Keep a clear route to a person for unclear requests, repeated recognition failures and complaints. Test the transfer when staff are busy and when no one answers. The customer should know what happens next, and the team should receive enough context to continue.

If you use your own cloned voice, document the voice owner’s permission, the intended use and who can change it. Make the automated nature of the service clear. Keep sensitive decisions outside the booking flow and obtain the necessary consent before messages or calls.

Measure a useful outcome

Track correctly completed bookings, duplicate records, successful transfers, correction work and the cost of a completed call. Include setup, integration maintenance, telephony and model usage. A cheaper minute is not a better deal if staff spend longer repairing the result.

Begin with one service and an agreed call volume. Compare the pilot against your current handling process, then expand only after it passes the real scenarios. Provider selection should follow the evidence from your business, language and systems.

Common questions