faster-whisper-gpu_skill

Overview

This skill provides high-performance local speech-to-text transcription using Faster Whisper with NVIDIA GPU acceleration. It enables private, offline transcription and subtitle generation across many audio formats and 99 supported languages. Runs on GPU for speed or CPU for compatibility, and supports multiple model sizes to balance accuracy and resource use.

How this skill works

The skill uses the Faster Whisper implementation to load Whisper-family models and run inference locally, leveraging CUDA when a compatible NVIDIA GPU is available. It accepts common audio formats, auto-detects language (or accepts an explicit language), and outputs plain text, SRT, JSON, or VTT with optional word-level timestamps. Command-line options and a Python API let you choose model size, device, precision, VAD settings, and transcription vs. translation tasks.

When to use it

• Transcribing audio locally when privacy is required and data must not leave your machine.
• Generating subtitles (SRT) with word-level timestamps for videos and podcasts.
• Batch-processing large audio collections without per-minute API costs.
• Quickly transcribing meetings or interviews using GPU acceleration for speed.
• Translating spoken audio into English using on-device translation mode.

Best practices

• Choose a model size that fits your GPU VRAM: tiny/base for low VRAM, medium/large for best accuracy.
• Use float16 or int8 compute types to reduce memory usage and increase throughput on GPU.
• Enable VAD to skip silence and speed up processing for long recordings.
• Monitor GPU memory (nvidia-smi) and fall back to CPU or smaller models if you see OOM errors.
• Provide an initial prompt or hotwords for domain-specific vocabulary to improve recognition.

Example use cases

• Create SRT subtitles for a podcast episode using --format srt and --word_timestamps.
• Transcribe multilingual meeting recordings with automatic language detection.
• Translate foreign-language interviews to English using --task translate.
• Run bulk offline transcription for archived audio without incurring cloud costs.
• Use CPU-only mode for lightweight, occasional transcriptions on machines without a GPU.