bitrix24_call_transcribe_max
Transcribe CRM calls with maximum quality using three local models. Returns raw transcript components for manual reconciliation into a clean transcript.
Instructions
MAX-quality transcription of a CRM call — the highest-fidelity pipeline, for when the basic bitrix24_call_transcribe isn't good enough. Runs THREE local models on the recording and returns the raw materials for you (the calling model) to reconcile into one clean transcript: • GigaAM v2 — Russian-native, never hallucinates, gets domain terms right (the reliable backbone) • Whisper large-v3 with condition_on_previous_text=False + domain hotwords — punctuation + proper nouns • pyannote diarization — speaker turns (who spoke when) Returns {whisper_text, gigaam_text, segments:[{start,speaker,text}], speakers, reconcile_hint}. YOU reconcile: keep Whisper for punctuation/proper-nouns, trust GigaAM where Whisper diverges into non-Russian garbage (hallucination), assign Менеджер/Клиент per speaker from content, fix diarization flips — then save with bitrix24_crm_timeline_note_save. Brand names are auto-normalised (V-LUX / вилюкс → Velux, etc.). REQUIRES: a heavy Python env (faster-whisper + gigaam + pyannote.audio + torch) at B24_MAX_PYTHON, and an HF token (env HF_TOKEN or B24_HF_TOKEN) whose account has accepted the pyannote gated-model terms. If any of that is missing the tool returns a clear, actionable error (error_type: missing_hf_token | missing_deps | model_not_approved) — fix that, then retry. Slower than basic (three models, no pooling).
Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| activityId | Yes | The call activity ID (a VOXIMPLANT_CALL activity) |