# Release Notes: Audio Analyzer This page tracks releases of the Audio Analyzer microservice. The most recent release is listed first; older entries are preserved for history. ## Version 2026.3.0 **Release Date:** TBD **New:** - Optional FunASR Paraformer models, VAD, and punctuation models can be downloaded from ModelScope when the CJK dependencies are installed. - Model snapshots can be provisioned through a shared Model Download service using `AUDIO_ANALYZER_MODEL_DOWNLOAD_URL`, with direct downloads available as a fallback. **Improved:** - Startup reuses valid model snapshots and provisions missing or incomplete snapshots from the configured source. - Updated the Level Zero package to version 1.34.0. ## Version 2026.2.0 **Release Date:** September 9, 2026 **New:** - OpenAI-compatible streaming transcription over Server-Sent Events (SSE), including incremental transcript events and completion signaling. - Realtime WebSocket transcription API for low-latency, bidirectional PCM16 audio streaming with server-side VAD and incremental transcription events. - VSS-compatible transcription APIs and response formats, including `/models` and `/transcriptions` compatibility routes. - OpenVINO NPU support for ASR, validated with Whisper Tiny and Whisper Base, with CPU, GPU, and NPU device selection. - OpenVINO-backed speaker diarization with GPU and NPU support. **Improved:** - Multi-speaker detection now independently analyzes Whisper segments, improving secondary-speaker detection in single-microphone kiosk scenarios. - Speaker enrollment persistence using `speaker_scope_id`, allowing speaker identity to persist across requests and service restarts. - Documented and standardized the supported ASR provider/device matrix across OpenAI, Whisper.cpp, and OpenVINO providers. ## Version 1.4.0 First release of the Audio Analyzer as a self-contained, OpenAI API-compatible speech-to-text microservice with optional voice sentiment analysis, built for edge deployment on Intel hardware. **Release Date:** June 17, 2026 **New:** - OpenAI-compatible transcription API (`POST /v1/audio/transcriptions`) and a streaming NDJSON variant (`/stream`). - Multi-backend ASR: `openai` (PyTorch Whisper), `openvino` (Intel-optimized), and `whispercpp` (CPU-only). - Full Whisper model family supported (`tiny` → `large`). - Optional voice sentiment analysis with session-level aggregation (`openvino` or `pytorch` provider). - FFmpeg-based preprocessing: chunking, silence detection, optional RNNoise denoising. - Session continuation via `session_id` (returned in `X-Session-ID`). - Health (`/health`) and ALSA device listing (`/devices`) endpoints. - New User Guide doc set, including: overview, get-started, how-it-works configuration, api-reference, and troubleshooting Markdown files, plus an architecture diagram and a restructured README.md. **Improved:** - OpenVINO CPU/GPU acceleration on Intel hardware; models warm-loaded once per process. - Layered config (`config.yaml`, env overrides via `AUDIO_ANALYZER__...`) and Docker Compose deployment on port `8010`. - Container now runs as a non-root user (UID 1000). **Known issues:** - The `prompt` form field is accepted for API compatibility but currently ignored. - Compatibility with the Video Search and Summarization sample application will be added in a subsequent release. ## Version 1.3.1 **Release Date:** April 2026 **New:** - Released as part of `release-2026.0.0`. - Supported features based on the requirements of the Video Search and Summarization sample application. Refer to that sample's release notes for details on this microservice at that version.