Audio Analyzer#

GitHub Readme

Audio Analyzer is a microservice that turns spoken audio into text and, optionally, into a high-level sentiment summary. It is designed to be dropped into voice-enabled applications (kiosks, assistants, call analytics, meeting notes) where a simple HTTP upload should return either a final transcript or a live stream of partial results.

Use Cases#

  • Conversational assistants and kiosks that need speech-to-text on the edge.

  • Post-call or meeting analytics where a session-level sentiment summary is useful alongside the transcript.

  • Batch transcription of recorded audio files.

  • Streaming transcription UIs that consume incremental NDJSON events as chunks complete.

Key Capabilities#

  • OpenAI-style transcription endpoint and a streaming NDJSON variant.

  • Multi-backend ASR (OpenAI Whisper, OpenVINO, with whisper.cpp planned).

  • Optional voice-sentiment analysis aggregated per session.

  • Session continuation so multiple uploads can extend the same conversation.

  • Runs on CPU; supports GPU acceleration on Intel hardware via OpenVINO.

Supported Models#

ASR (speech-to-text):

  • Whisper family — whisper-tiny, whisper-base, whisper-small, whisper-medium, whisper-large — selectable via models.asr.name.

  • Backends: openai (PyTorch), openvino (Intel-optimized).

Sentiment (optional, voice-based):

  • Default: speechbrain/emotion-recognition-wav2vec2-IEMOCAP.

  • Any compatible Hugging Face model can be configured via sentiment.model, served through the openvino or pytorch provider.

Next Steps#