Voice

Overview#

Anar Voice is an Arabic speech intelligence platform designed for GCC government organizations. It provides real-time speech-to-text, dialect detection across 5 Arabic dialect families, speaker diarization, and conversational voice agents — all powered by Groq Whisper and Arabic LLMs.

Government call centers, citizen service desks, and meeting rooms generate enormous volumes of Arabic speech data. Voice transforms that audio into structured, searchable, and actionable text with sub-second latency.

Key Capabilities#

Real-Time Transcription#

Powered by Whisper Large V3 via Groq's inference infrastructure, Voice transcribes Arabic and English audio with high accuracy. Both batch file uploads and real-time WebSocket streaming are supported, enabling live transcription for call centers and meeting rooms.

Arabic Dialect Detection#

Voice identifies 5 Arabic dialect families — Gulf, Egyptian, Levantine, Maghreb, and Modern Standard Arabic — enabling downstream systems to route, tag, and analyze speech by regional origin.

Conversational Voice Agents#

Built-in voice agent capability combines speech-to-text with LLM reasoning and conversation memory. Users speak naturally, and the agent transcribes, understands, and responds in context — ideal for automated citizen inquiry lines.

Speaker Diarization#

Multi-speaker audio is segmented by speaker, enabling accurate attribution in meetings, interviews, and call recordings.

Dialect Analysis Dashboard#

The dashboard provides dialect distribution charts, language breakdowns, and transcription filters by dialect family, duration, and language -- enabling operational insights into the linguistic patterns of processed audio.

Export to Chat#

Transcription results can be exported directly to Anar Chat for follow-up Q&A, enabling seamless voice-to-text-to-conversation workflows.

Guard Telemetry#

When connected to Anar Guard, Voice reports LLM and STT call telemetry for centralized governance monitoring.

Architecture#

Voice follows a two-tier architecture with a FastAPI backend on port 8003 and a Next.js dashboard on port 3003.

The backend integrates with Groq for both Whisper transcription and LLM chat (llama-3.3-70b for English, allam-2-7b for Arabic). Transcription history and analytics are persisted in SQLite or PostgreSQL.

Quick Start#

cd voice/backend
uv sync
uv run uvicorn anar_voice.main:app --reload --port 8003

Test a transcription:

curl -X POST http://localhost:8003/api/v1/transcribe \
  -F "file=@meeting.mp3" \
  -G -d "language=ar"

Dashboard#

The Voice dashboard provides three views:

• Live Transcription — Record directly from the browser or upload audio files for real-time transcription
• Transcription History — Browse, search, and filter past transcriptions with full text and segment data
• Analytics — Call statistics, language distribution charts, and duration breakdowns

Observability#

Voice is instrumented with OpenTelemetry via anar_shared.setup_otel(). When OTEL_EXPORTER_OTLP_ENDPOINT is set, traces, logs, and metrics are exported to the OTel Collector.

MCP Server#

Voice exposes a Model Context Protocol server at /mcp via FastApiMCP, enabling AI assistants to trigger transcriptions, query history, and run dialect detection programmatically.

Gateway Integration#

When GATEWAY_URL is set, Voice routes LLM calls through Anar Gateway for unified cost tracking and safety scanning.

Test Suite#

260 tests cover transcription, dialect detection and analysis, voice agents, speaker diarization, Chat export, transcription filters, telemetry, and API endpoints.

cd voice/backend && uv run pytest

Next Steps#

• Transcription — Whisper models, batch processing, and segment-level output
• Dialect Detection — How dialect classification works across 5 Arabic families
• Voice Agents — Building conversational agents with WebSocket streaming
• API Reference — Full endpoint documentation