stt-service

To transcribe Hungarian speech to text in a multi-tenant microservice, you deploy a shared Whisper model via FastAPI, sending multipart/form-data audio uploads to the /transcribe endpoint to receive real-time transcripts.

Yes, you can use GPU acceleration for Whisper speech-to-text inference in FastAPI. The service supports GPU/CPU selection during startup-time model loading to accelerate per-request transcription of audio files.

You configure the Whisper model and language for the speech-to-text API by setting the STT_MODEL_ID and STT_LANGUAGE environment variables at startup, and you can also override the language per request via the /transcribe endpoint.

Yes, the speech-to-text FastAPI service provides the /health endpoint for status checks and the /transcribe endpoint to accept multipart/form-data audio uploads, returning the transcript, detected language, and timing metadata.

The best way to handle Hungarian audio transcription for multiple tenants is deploying a centralized speech-to-text microservice, which simplifies deployment and maintenance by sharing a single Whisper model across tenant applications.

A limitation of using a shared Whisper service for real-time call transcription is that model loading occurs at startup, meaning per-request transcription depends on the initial STT_MODEL_ID configuration and available GPU or CPU resources.