Voice · on the roadmap
Calls in, answers out
A caller speaks. Their words become text, run through the same pipeline as every other channel, and come back as speech, as a text message, or as a person on the line. This page is what we are building, in what order, and what is genuinely hard about it.
Why it fits
A call is just another message, once it is words
Broa separates how a message arrived from what it says. Everything downstream — classification, retrieval, generation, approval — reads only the second part. So a transcribed call needs no new pipeline, no second set of tone rules and no duplicate catalogue. Voice is an adapter at the front door and a voice at the back.
Which is why the hard part is not speech recognition
Turning speech into text is a bought component and it works well. The difficulty is that a person on the phone will not wait. An email can take thirty seconds to answer and nobody notices; a caller needs a response in about a second or the silence becomes unbearable. Broa was built for messages that queue patiently. Making the whole path answer that fast is the actual project.
The plan
Three steps, in this order
The first two carry no latency risk at all. Shipping them together with the third would mean shipping the hard one having learned nothing from the easy ones.
Voicemail & after hours
A call that arrives at 7pm is transcribed, answered from your catalogue and order records, and sent back by email or text before anyone opens the office. Nobody is waiting on the line, so the whole latency problem disappears — and it solves a real one on its own: calls that land out of hours and get dealt with the next morning.
Answered by text
A live caller is offered the answer as a text message rather than spoken aloud. It dodges the speed problem again, and it is often the better outcome anyway: the customer gets a tracking number and a price they can read back later, and both sides end up with a written record.
Spoken conversation
Real-time spoken answers to a caller on the line. This ships when it is genuinely fast enough and not before — with a target of a first response inside 1.2 seconds, a holding phrase rather than silence when a lookup runs long, and interruption handled so a caller can talk over it.
What does not change
The rules are the same on the phone
- Grounded, or transferred
- A spoken question Broa cannot answer from your data goes to a person — by transfer, a callback offer, or a text follow-up. It is not guessed at because a caller is waiting.
- Validation still applies
- A generated answer that fails validation is not spoken. There is no policy setting that releases something the written channels would have blocked.
- Approval works differently, and we say so
- You cannot hold a live call in a review queue. On the phone you are either running under a policy you configured, or a person is on the line being assisted. Voicemail, where nobody waits, uses the ordinary review queue like every other channel.
- Low confidence is treated as low confidence
- A noisy line or an unfamiliar accent lowers recognition confidence, and a message Broa is unsure it heard correctly is not answered automatically at any autonomy setting.
- A handover carries the context
- Transfer to a person brings the transcript so far, the identified intent and the facts already retrieved. Your colleague does not start from "how can I help?"
Recording & consent
Stricter than anything email asked of you
Consent is per number and per jurisdiction
Rules differ by country and by US state, including where both parties must agree. Announcements play before anything is retained, and the announcement itself is recorded as having happened.
Declining recording does not end the call
Where a caller declines and consent is required, the call proceeds without audio being kept. Whether the transcript may still be retained is configuration, and the distinction is made explicit rather than assumed.
Audio and transcript expire separately
Audio is the larger liability and can be set to expire long before the record that a call took place. Spoken card numbers are redacted from both before anything is stored durably.
Listening is an audited read
Playing back a recording is recorded the same way opening a message is. Who listened, and when, is part of your trail.
Want voice sooner?
Tell us which of the three steps would matter to your business and roughly how many calls a week you take. It genuinely affects the order we build in.