Call transcripts

A transcript is not a recording#

The two are often spoken of together and they are different artefacts with different costs, different risks and different availability. A recording is audio: it captures tone, background, accent and hesitation, it is large, and in most jurisdictions it carries notification duties. A transcript is text: it captures words and loses everything around them, it is small enough to search, and it is what a machine can act on.

That asymmetry is deliberate. The text is what the engine needs to write a summary, capture a lead and honour a promise; the audio adds evidence a person might want and a compliance obligation nobody asked for.

Where the words come from on each engine#

The second row is the one that surprises people. On a speech-to-speech model the reply is generated from audio directly, so the transcript is a description of the conversation rather than the input to it. A word written oddly in the transcript was not necessarily heard oddly, and a reply that reads wrong may have sounded right.

It also means transcription of a language written in another script is normal rather than broken. A Telugu reply transcribed in Latin letters is still Telugu — romanised_script in the language tracker exists exactly so that a romanised line is not mistaken for an English one.

Does Connect use call transcripts?#

Used, on every call, on both engines. The transcript is the input to five separate things, and knowing which is which explains most questions about why a call produced what it did.

• The call summary, written at the end of the call by a model reading the turns.
• Lead capture, which reads what the caller said about themselves and their need.
• Follow-up integrity — schedule_follow_up is given the caller's last lines and the voice's own, and record_event refuses a promised time that shares no time word with the caller's latest time-bearing line.
• The quality review, whose deterministic findings include repetition, stock phrases, wrong script and an unheard caller.
• Guardrails after the fact: the realtime worker posts every sentence it speaks to the engine, which re-checks it rather than trusting the model's own restraint.

There is a floor on the first of those. A call with fewer than two caller lines and fewer than five caller words is never sent to the summariser. That rule exists because ten one-turn rows once carried an invented afternoon at a sales desk: a summariser given almost nothing writes something plausible rather than nothing.

The clock a transcript is not#

Reading response time out of transcript timestamps is the single most common way to measure a voice agent wrongly, and the numbers it produces are not slightly wrong — they are wrong in the flattering direction. A transcript row is written when a sentence is *complete*. The caller stopped speaking long before that, and the reply began before its own row existed.

The first real call reviewed here showed 8 ms beside replies the caller had waited seconds for. Latency is now taken on the wire instead: the worker's TurnClock times the caller's last word — the moment user state leaves *speaking* — to the first audio back, when agent state enters *speaking*, and posts a turn_timing event per reply.

What is kept out of the text on purpose#

Supervisor guidance

An instruction a colleague gives during a live call goes into the system prompt and nowhere else — not into the spoken reply, not into call_turns, not onto the wire.

Rejected names

Every door a name passes through is guarded. A form of address is not a name, and a name-keyed fact that is a suffix, a pronoun or 'the caller' is refused rather than stored.

Control actions

Pause, take over, mute and hang-up ride the guidance log and are never spoken, so they leave no line in the conversation.

The read-back mechanism works the other way: when a phone number, an email address or a code appears in the caller's words, a note records what was heard and the next two lines are checked for confirmation. That is a deliberate use of the transcript as evidence rather than as memory.

Questions#

Related#

• Call transcripts
• Call summaries
• Speech recognition
• Call quality measurement
• Reply latency on a call
• Calls that were never a conversation