One speech engine for every SAIG app. The exact words, always. On top: dictation that reads like you typed it, and call intelligence that explains the reactions in the room, including the complaint nobody picked up.
Say it once, with the ums and the corrections. SAIG Voice writes what you meant, in the style you ask for: clean, professional, concise, a message, an email, notes, or a prompt for an AI.
For relationship managers, sales, support and advisers, and for compliance QA. SAIG Voice reads a call the way a senior, empathetic listener would, and says what the agent missed: subtle complaints, confusion, frustration, objections, buying signals, vulnerability cues under the UK FCA's Consumer Duty, commitments, and the risk of losing the client. Every point cites the exact words.
The verbatim words, untouched, with the reaction explained inline where it matters.
What the client signalled, what the agent did, and what a good response would have been.
Health, bereavement, job loss, money worries, capability: flagged, with whether they were handled well.
Two to four coaching points per call, every commitment either side made, and an overall trajectory, satisfaction score and risk.
Every finding says how the adviser responded (well, partly, not) and whether it is a defect. Only defects count, so a moment handled well never reads as a fault.
Send your question list and each finding is mapped to the questions it bears on. ID&V is read as a structured check: what was asked and whether it was completed.
Speakers come back as agent, client or third party, and words the engine was unsure of are marked so nobody is marked down for a mishearing.
A word-error-rate check against human-corrected transcripts compares engines and settings on numbers, and every analysis carries a version so results can be re-checked when it changes.
How tone is judged, stated on every response: the words, plus timing measured from the audio (pace, pauses inside a turn, silence before answering, interruptions, hesitations). Not voice pitch or loudness.
Call insights on the ended page of every recorded meeting and "Tidy notes" in the call.
Consumer Duty QA on recorded client calls: missed signals and vulnerability, pseudonymised in the UK first.
Dictation in the assistant, polished before it is sent.
Hisaab, SKYF, Ilm, Hello and the rest call the same engine with their own token.
Call the front with your app's own token (or a person's SAIG sign-in). Everything is metered to your app by SAIG AI; the verbatim is always returned.
POST https://voice.skylite.group/api/voice?op=transcribe
headers x-saig-service-token: <your app token> (or Authorization: Bearer <SAIG sign-in>)
body { "audioUrl": "https://…/call.mp4", // audio or video file; or "audioBase64" + "mime" (≤ 8 MB)
"mode": "analyse", // dictate · note · converse · analyse
"insights": { "context": "RM call with a private-banking client", "focus": "fee changes" },
"polish": { "style": "notes" } }
→ { "verbatim": "…untouched…", "utterances": [...], "polished": { "text": "…" },
"insights": { "summary", "overall": { "trajectory", "satisfaction_0_10", "risk" },
"moments": [{ "t", "kind", "evidence", "meaning", "acknowledged", "better_response", "severity" }],
"missed_signals": [{ "t", "signal", "what_agent_did", "should_have" }],
"vulnerability": [...], "actions": [...], "coaching": [...], "annotated_transcript", "tone_basis" } }| Mode | For | Returns |
|---|---|---|
dictate | one speaker, short-form | verbatim + polished text |
note | dictation into tidy notes | verbatim + bullet notes |
converse | multi-speaker, diarised | verbatim with speakers |
analyse | a call or meeting | verbatim + insights |
Also: POST ?op=vocabulary { terms } for your app's names and jargon · POST ?op=wer_eval for accuracy · text-only utterances when you already hold a transcript · GET ?op=usage · the colleague card at /api/ext · health. Engine: console.skylite.group/api/ai?op=transcribe; the full contract is in the estate's SAIG Voice spec §9.
converse or analyse.