August 18, 2026
Why We Show You Every Call's Transcript
A call transcript phone system should show you the underlying text, not only a tidy summary of it. Signal displays the full transcript alongside every AI call summary and voicemail transcription, because a summary is a claim about a call and a transcript is evidence. We also ship both features off by default, for a reason worth stating plainly.
Start with the caveat, because it decides whether you use this at all
Signal's transcription and summary features send call audio and text to third-party AI providers — OpenAI for Whisper transcription, Anthropic for summaries. We do not have a Business Associate Agreement with either provider.
That is why both features are off by default on every client account, and why turning them on is a deliberate act by a named person rather than something that happens because you signed up.
If your practice handles protected health information, leave AI transcription and AI call summaries off until a BAA exists. We would rather tell you that in the second paragraph of a blog post than have you discover it during an audit. The rest of the phone system — routing, queues, voicemail delivery, reporting — works exactly the same with those toggles off.
Nothing about a phone system makes an organization HIPAA compliant, ours included. HIPAA and your phone system covers what actually applies and what is marketing.
What does Signal actually transcribe?
Two things, both after the fact.
Voicemail messages. When transcription is enabled, Signal runs Whisper over the recorded message and delivers the text with the voicemail notification, so you read it in your inbox instead of dialing in.
Recorded inbound calls. When call recording is enabled, Signal transcribes the recording and produces a summary from it, with action items, a sentiment classification, and a red-flag indicator on calls where the caller had a bad experience.
Neither is live. Signal does not transcribe a call in progress, does not display running text to an agent, and does not act on a call while it is happening. Everything here operates on a recording that already exists.
Why show the transcript and not just the summary?
Because a summary is a model's compressed reading of a conversation, and you occasionally need to check the reading.
We built the summary to be the thing you use daily: three or four lines, action items pulled out, sentiment noted. That is the right unit for a manager reviewing yesterday. But summaries flatten things. A caller who said "I guess that's fine" is not the same as a caller who said "that's fine," and a summary will usually render both as agreement.
So the transcript sits one click away, always, on every summarized call. Three situations make it non-negotiable:
- Disputes. Somebody says they were quoted a different price. The summary will not settle it. The transcript and the recording will.
- Red flags. When Signal marks a call as a bad experience, the first question is always "what actually happened?" A red flag you cannot inspect is an accusation, not information.
- Trust in the tool. Staff who can check the model's work stop treating it as a black box. Staff who cannot check it either over-trust it or ignore it entirely, and both are worse than the transcript.
We also think a vendor that shows you only its own summary is asking for more trust than it has earned. Show the source.
What has to be true before any of this works?
Call recording, and call recording has a legal requirement attached to it.
Signal requires a consent announcement on recorded calls. You upload the audio, Signal plays it to the caller before recording starts, and recording is inbound-only by design. This is not a courtesy setting. Federal law permits one-party consent, but a substantial number of states require all parties to consent, and a recorded call without notice in one of those states is a legal problem rather than a compliance checkbox.
Record the announcement in the same voice as your other prompts and keep it short. "This call may be recorded for quality purposes" is sufficient and understood. The recordings library covers uploading prompts and greetings; call recording laws by state covers which states change the calculus.
If you cannot live with the announcement, you cannot have recordings, and without recordings there are no call transcripts or summaries. That order is deliberate.
What do you get back from a transcript you can read?
The honest answer is triage speed and pattern visibility, not magic.
A transcript is scannable in a way that audio is not. You can read a five-minute call in about forty seconds, search it for a policy number, and forward the relevant paragraph to the person who needs it. Audio makes you sit through the whole thing at the speed it was spoken.
Across many calls, the summaries become the useful layer. Action items surface commitments staff made and forgot. Sentiment classification, applied consistently, turns "I think Tuesdays are rough" into something you can look at. Red-flag detection surfaces the call that went badly today, rather than the review that appears next week.
What you see | Where it comes from | What it is good for |
|---|---|---|
Full transcript | Whisper, run on the recording | Disputes, exact wording, search |
Summary with action items | Claude, run on the transcript | Daily review, follow-up |
Sentiment | Same model pass as the summary | Trends across many calls |
Red flag with reason | Same model pass as the summary | Catching a bad call the same day |
Signal generates the summary, sentiment, and red-flag classification in a single model pass per recording, which is why they always agree with each other.
What we deliberately do not do
We do not score your staff. Sentiment and red flags describe the caller's experience, not an employee's performance rating, and we have no interest in shipping a number that a manager treats as a grade.
We do not transcribe outbound calls. Recording is inbound-only because the feature exists to understand what customers experience.
We do not send anything to a model when the toggles are off. An account with AI features disabled produces recordings and nothing else — no transcripts, no summaries, no red flags, no data leaving Signal for a model provider.
And we do not make transcripts the default. A phone system that quietly starts transcribing everything the day you sign up is a phone system that made a legal decision on your behalf.
Frequently asked questions
Does Signal transcribe phone calls? Yes, when you enable it. Signal transcribes voicemail messages and recorded inbound calls using Whisper, then generates a summary with action items, sentiment, and red-flag detection from the transcript. Both features are off by default per client and require call recording with a consent announcement for the call-transcript half.
Is call transcription HIPAA compliant? No, and Signal does not claim it is. Transcription and summaries send audio and text to OpenAI and Anthropic, and Signal has no BAA with either provider. A practice handling PHI should leave both features off pending a BAA. The rest of the phone system operates normally with them disabled.
Do I need consent to record and transcribe calls? Yes. Signal requires a consent announcement that plays before recording begins. Federal law allows one-party consent, but many states require all parties to consent, so the announcement is legally load-bearing rather than a formality. You upload the audio and Signal plays it on every recorded call.
Can I turn call transcription off? Yes, and it starts off. AI transcription and AI call summaries are per-client toggles set to disabled by default. Turning them off stops all model processing immediately; recordings continue if call recording remains enabled, and existing transcripts stay in the portal until deleted.
If you want the summaries but the BAA question is unresolved on your side, that is a conversation worth having before you sign anything. Tell us what you handle and we will tell you plainly which features you should leave off.