Model Card · Sentinel Voice-Note Transcription (OpenAI Whisper-1)

Altara Sentinel allows customers, employees and managed-services analysts to upload voice notes they have received as part of a suspected scam — for example a WhatsApp voice message claiming to be from a bank. The recording is transcribed to text before NAVI analyses it. This card describes the transcription model, how it is used, what is stored, and the limitations.

1. Purpose & Intended Use

The transcription model exists to convert a single, user-supplied audio sample into a text transcript so that:

1. NAVI can analyse the transcript content for scam indicators (urgency, impersonation cues, payment requests, link mentions).
2. The case file shown to the human Sentinel analyst contains a readable version of the audio for evidentiary review.
3. The Data-Rights Bundle exported to the user contains the transcript so they can verify what was processed.

Intended users

• Sentinel public submitters uploading suspicious voice notes for trust-score analysis.
• Sentinel managed-services analysts triaging escalated cases.

Out-of-scope

• Whisper-1 is not used for speaker identification, voice biometrics, deepfake detection, sentiment analysis, or any kind of profiling beyond converting audio to text.
• It is not used to transcribe call-centre recordings or any live audio stream.
• It is not used on audio the submitter has not chosen to upload to Altara.

2. Inputs

• A single audio file uploaded by the submitter via the Sentinel submission flow.
• Supported formats follow Whisper-1's accepted list (mp3, m4a, wav, ogg, webm, mp4 audio container, mpga, mpeg).
• File-size guarded both client-side and at the FastAPI boundary.

No additional metadata is sent to the provider beyond the audio bytes themselves and the request to transcribe.

3. Outputs

• A plain-text transcript of the audio. No timestamps, diarisation or speaker labels are requested.
• The transcript is stored in the Sentinel case record alongside a hash of the original audio file.

4. Training Data & Lineage

OpenAI Whisper was trained on a large multilingual speech corpus that Altara does not control and cannot independently verify. Altara has not contributed audio to the training of Whisper. The model is used in inference-only mode; no submitted audio is used to fine-tune the base model.

For full base-model lineage see OpenAI's Whisper paper and the OpenAI Audio API documentation.

5. Performance & Known Quality Bounds

• English — Altara's internal sample testing shows usable transcripts on clean and moderately noisy WhatsApp voice notes. Errors increase with low bit-rate audio, accented speech and background noise.
• South African accents — usable but error-prone on regional pronunciation; analyst review is mandatory.
• isiZulu / Afrikaans — partial. Whisper-1's accuracy on local Southern African languages varies by speaker; Altara plans to add a second-pass model for these languages (see Sentinel Phase B in the product roadmap).
• Audio with multiple speakers — Whisper-1 produces a single interleaved transcript without speaker labels; analysts must reconcile speakers manually if relevant.

Because the transcript is always reviewed by a human Sentinel analyst before any externally-facing action, ASR errors do not flow unchallenged into a decision.

6. Human-in-the-Loop Checkpoints

1. Analyst review — every transcript is shown to a human Sentinel analyst alongside the original waveform link. The analyst must explicitly accept the transcript before any case-file action.
2. Submitter visibility — the submitter can view the transcript in their case view and request correction or deletion via the Data-Rights Bundle.

7. Data Flow & Residency

[ Submitter device ]
       │  (HTTPS upload)
       ▼
[ Altara FastAPI backend · audio stored encrypted at rest ]
       │  (Universal Emergent LLM Key · HTTPS)
       ▼
[ Emergent gateway ]
       │
       ▼
[ OpenAI Whisper-1 — synchronous transcription · no persistent storage at provider ]
       │
       ▼
[ Transcript stored on the Altara case record · linked to the original audio hash ]

• OpenAI publishes that API audio is not retained for training purposes by default. Altara relies on that contractual position.
• The original audio file remains on Altara-managed storage; transcripts are stored alongside the case.
• Submitters can request deletion of both the audio and the transcript via hello@altaracore.ai (subject [ALTARA-DATA-RIGHTS]) or via the Data-Rights Bundle download → delete flow.

8. Known Limitations & Failure Modes

• Word-error rate scales with audio quality, accent, code-switching and background noise.
• Mis-transcription of named entities (bank names, scam URLs) — the analyst is responsible for catching these before any escalation.
• No deepfake / synthetic-audio detection — Whisper-1 transcribes whatever it receives. Detecting synthetic audio is a separate Sentinel Phase B initiative.
• Privacy spillage risk — if the submitter inadvertently uploads audio containing sensitive information about a third party, the audio still transits the transcription path. Altara mitigates with a pre-submission warning and a quick-delete option after submission.
• Language coverage — accuracy outside the well-supported language list is poor; do not rely on transcripts in those languages without analyst confirmation.

9. Monitoring, Incident Handling & Drift

• Every transcription call is logged with prompt version, model ID, latency and bucketised quality signal (analyst-accepted, analyst-corrected, analyst-rejected).
• Quality-signal trend is reviewed monthly by the Altara Sentinel team.
• OpenAI Whisper-1 model upgrades are tracked via the Emergent gateway release notes. The model ID is pinned in Altara's code so silent upgrades are impossible without a deliberate code change.
• Kill switch — Sentinel transcription can be disabled by feature flag without redeploy.

10. Frameworks This Card Maps To

11. Versioning & Change Log

A new card version is published when the underlying model changes, when accepted languages are expanded, or when the data-flow / residency posture changes.

12. Contact & Reporting

• Card owner / questions — hello@altaracore.ai
• Responsible AI / incident reporting — hello@altaracore.ai · subject [SENTINEL-TRANSCRIBE-INCIDENT]
• Right-to-deletion — hello@altaracore.ai · subject [ALTARA-DATA-RIGHTS]

Altara Sentinel is a module of Altara Core, a division of Navigate Group (Pty) Ltd · Reg No 2016/343423/07. This document is published as a public AI transparency artefact. © Navigate Group (Pty) Ltd — all rights reserved.