FasterWhisper

Speech transcription engine used by the Speaches runtime.

Agentic Friendly

Component Category

Inference / speech-to-text

Component Description

FasterWhisper is an optimized implementation of Whisper for efficient speech transcription workloads. BullSequana AI runs it inside Speaches rather than as an independently deployed component: in the backend-owned Speaches service of the BSQAI API release, and in the Speaches serving runtimes on which speech models registered in MLflow deploy as KServe InferenceService workloads.

Why It Is Used

In BullSequana AI Foundation, FasterWhisper supports audio transcription use cases with lower latency and lower resource consumption than a baseline Whisper deployment.

Learn More

Deployment notes

FasterWhisper runs inside the Speaches service. Speech-to-text is exposed through the authenticated BSQAI API, which supports batch transcription and real-time WebSocket sessions. Operators select CPU or GPU execution through the AI service configuration; the selected profile's model list decides which Whisper models are installed. The defaults are Systran/faster-whisper-base on CPU and deepdml/faster-whisper-large-v3-turbo-ct2 on GPU. Clients choose one installed model per request.

Interacts With

  • BSQAI API, which exposes batch and real-time speech-to-text operations.
  • Speaches, which hosts FasterWhisper and the platform's speech-generation engines.
  • KServe, which serves registered speech models on the Speaches serving runtimes.
  • AI Web Portal, which uses speech services for voice dictation.

On this page