Transcribe Audio (Faster Whisper)

Summary: CTranslate2-accelerated Whisper with built-in Silero VAD filtering, 4x faster than vanilla Whisper with configurable hallucination suppression.

Use This When

• You want Whisper-quality transcription with significantly better performance
• You need built-in VAD filtering to suppress hallucinations on silence
• You need flexible model size selection from tiny to large-v3-turbo
• Multilingual support across 100+ languages is required

What It Does

• Converts audio frames to text using Whisper models via CTranslate2 backend
• Built-in Silero VAD pre-filters silent segments before they reach the decoder
• Supports INT8 and FP16 quantization for speed/memory tradeoffs
• Configurable no_speech_threshold and beam_size for accuracy tuning
• 4x faster than vanilla Whisper at equivalent accuracy

Works Best With

• Drop-in replacement for the standard Transcribe Audio component with better performance
• Pair with detect-voice-activity for additional silence pre-filtering
• Voice pipelines where Whisper compatibility and multilingual support matter

Caveats

• Still hallucinates on silence (Whisper architecture limitation) but VAD filter mitigates most cases
• Requires CUDA 12 + cuDNN 9 for GPU inference
• Autoregressive decoder is inherently slower than CTC-based alternatives for short utterances
• Model auto-downloads from HuggingFace on first run (size varies by model)