September 23, 2026. Twilio posted a Voice changelog entry on September 22 that most teams will skim and should not. Starting October 22, 2026, Twilio Real Time Transcriptions will automatically start a transcription session on a second speech provider when the primary provider does not connect. The entry states that failover is enabled for all existing Real Time Transcriptions accounts on that date. It is optional, but it is opt out rather than opt in, and the opt out is per sub account in Console. If you run AI voice agents for clients, you have one month to decide what you want on each sub account.
What Twilio is actually changing
- Until now a transcription session depended on one speech provider being reachable, and a session that could not connect returned an error. From October 22 Twilio retries the alternate provider instead, so a session starts on Deepgram if Google does not connect and on Google if Deepgram does not connect.
- Failover triggers at session startup only, never mid session, and it stays inside your existing region, so a session processed in Europe fails over to the alternate provider in Europe.
- Twilio is also launching an Auto mode on the same date, which becomes the default TranscriptionEngine selection for new Real Time Transcriptions customers. Auto mode lets Twilio pick the provider per session on criteria including word error rate for the language and availability.
- Existing customers keep their configured transcription provider. Auto mode is opt in for them, by setting transcriptionEngine to auto on the Transcription TwiML noun or on Calls Transcriptions API requests.
- Sessions that use the language value multi are unchanged and continue to run only on Deepgram's Nova 3 multilingual model.
The line that matters is about what gets dropped
Twilio is explicit about the trade it is making. If the alternate provider does not support profanity filtering, custom hints, an exact language variant, or provider word timing data, Twilio drops or downgrades that setting and returns a working transcript instead of an error. The setting applies as normal on the next session with the usual provider.
Read that as an operator rather than as a developer. Today a provider outage produces an error, and an error is loud. Your monitoring catches it, the call is retried or routed to a human, and somebody knows. From October 22 the same outage produces a transcript. A slightly worse transcript, with your custom hints missing, delivered successfully, with nothing in your logs flagging that anything was different.
Why custom hints and word timings are not cosmetic
Custom hints are how an AI phone agent learns vocabulary that general speech models get wrong. For the home services clients we build for, that is the difference between a caller saying condenser, Rheem, or a street name the model has never seen, and the agent hearing something else and booking the wrong job. Agencies building AI receptionists on our AI calling for home services stack typically load dozens of hints per client: trade terms, brand names, neighbourhood names, the firm's own service list.
Provider word timing data is what barge in and interruption handling are built on. An agent that cannot tell when the caller started speaking will either talk over them or leave dead air. Lose the timings silently and the call quality drops in a way that shows up in booking rates weeks later, not in an alert that night.
The cost side is unchanged by this, but if you have not priced the transcription leg of a voice agent recently, the figures are in our breakdowns of AI transcription cost per hour and what an AI phone call costs per minute.
What it means for operators
Do three things before October 22.
First, decide per sub account rather than globally. For a client whose agent only needs to capture a name, a number and a reason for calling, failover is strictly better than an error and you should leave it on. For a client whose agent depends on custom hints or word timings, an error you can see may be worth more than a transcript you cannot trust, and the Console opt out is there for exactly that case.
Second, start logging which provider served each session now, while the answer is always the same one. After October 22 that field is the only way to tell a normal call from a failed over call, and you cannot backfill it.
Third, alert on the downgrade rather than on the outage. If a session comes back without the hints or timings you asked for, treat that as an incident with a quality consequence, not as a successful call. That is the same discipline that keeps messaging builds honest, which we covered in the 10DLC daily limit explainer.
Sources: Twilio Voice changelog, Introducing Provider Failover and Auto mode Provider selection for Real Time Transcriptions on October 22, 2026, dated September 22, 2026, cross checked against the Twilio TwiML documentation for Real Time Transcriptions and the Calls Transcriptions API reference linked from that entry.
Frequently Asked Questions
October 22, 2026. Twilio's changelog entry of September 22 states that failover is enabled on that date for all existing Real Time Transcriptions accounts.
Yes. Twilio describes failover as optional and says it can be opted out of per sub account in Console, which means the decision can be made client by client rather than once for the whole account.
No. Failover triggers only at session startup when the primary provider does not connect, never mid session, and Twilio says the normal setting applies again on the next session with the usual provider.
Twilio states that if the alternate provider does not support profanity filtering, custom hints, an exact language variant, or provider word timing data, it drops or downgrades that setting and returns a working transcript rather than an error.