Opal (Deepslate)

Current status: disabled in DEV. French tests did not produce a satisfactory accent. The model is no longer offered for new bots and model availability blocks new calls to its test bot. The integration, previews and existing configuration are retained for possible reactivation after a suitable French voice is validated. The details below describe the retained integration, not an active service.

Speech detection and accent

If the bot speaks but does not react to a caller who is audible in the recording, check the custom turn-detection threshold. A threshold suitable for another provider can be too strict for Opal. Start testing at 0.40, retaining 300 ms of end-of-speech silence. Lower thresholds also increase sensitivity to noise; validate using real telephone conditions. Threshold and silence duration are separate settings.

The equivalent API setting is "turn_detection": {"override_enabled": true, "threshold": 0.4} in PATCH /api/v1/bots/{id}/voice. Parameters without an override continue to inherit system defaults. A transcript added after hangup may come from final transcription and does not prove the live model heard that utterance.

Hosted voices have no dedicated accent setting in this integration. Deepslate supports reference voice cloning and ElevenLabs output, but neither is currently offered by the BeAI Opal voice selector. A French reference voice is a separate voice configuration, not a prompt adjustment.

Opal handles live audio conversations using Deepslate-hosted voices and the same authorized BeAI skills as other voice models.

Set up and preview

In Bots → Edit → Voice, select Opal 2.1 (Deepslate) and Alma (default), Hannah, Lisa, Marina or Jess. Preview samples follow the bot language: French, Belgian French, Canadian French, English, Dutch, Spanish or Catalan. Samples use fixed demonstration text, not private bot instructions. A language sample does not guarantee a regional accent. Previewing does not place a call or modify the bot; unavailable samples fall back to the primary language and French.

Speech speed is fixed at 1. Server VAD uses automatic responses and interruption, inheriting the system default 300 ms of silence unless overridden per bot. Semantic VAD and automatic idle prompts are unavailable. Model processing adds further latency. Prompts, greetings, telephony, recording and history follow the usual setup. Tool results automatically resume the response. The final report after hangup is silent.

An active model rate is required in the organisation hierarchy. Administrators manage rates through the standard pricing workflow. Platform administrators configure the provider credentials, which are never exposed to bot users.

Public API

The GUI and API enforce the same bot editing permissions and organisation scope. Use PATCH /api/v1/bots/{id}/voice:

{"ai_model":"opal-v2.1","voice":"de56bbd1-3ea7-4543-a17c-504664d9116b","speech_speed":1.0,"turn_detection":{"override_enabled":false}}

Reading the bot returns these settings. The OpenAPI schema lists all five voice IDs. Both surfaces reject incompatible voices, unsupported speed and unsupported VAD settings. Another organisation's bots remain inaccessible.

Usage and failures

BeAI credits follow the organisation's configured model rate. Provider usage is a BeAI session duration estimate, rounded up to six-second increments, not provider-reported tokens or a certified invoice. Deepslate's portal lists EUR 0.10/minute for Opal V2.1 on September 16, 2026. A supplier cost is shown only when a rate is configured in the platform currency; EUR is never silently interpreted as USD.

Missing credentials, provider errors or exhausted provider credits can prevent or interrupt a call. Administrators can inspect the call status and history. A failed preview stops playback; retry or select another voice.

See the official WebSocket protocol and realtime reference.