Skip to main content

Private audio transcription

Transcribe and translate audio without sending data to cloud services. Ideal for sensitive content.

VRAM mínima: 8 GB Para: Journalists, researchers, healthcare professionals GPU referencia: ~$270–380 (used RTX 3060)

GPUs compatibles con este setup

Este escenario necesita al menos 8 GB de VRAM. Estas GPUs pueden ejecutarlo:

Precios y disponibilidad pueden variar. Ver todas las GPUs NVIDIA →

Por qué este setup

Este escenario está diseñado para journalists, researchers, healthcare professionals. La RTX 3060 12 GB ofrece el equilibrio óptimo entre capacidad (8 GB de VRAM mínimos), disponibilidad en el mercado y coste relativo para los casos de uso de este escenario.

Con 8 GB de VRAM puedes cargar los modelos recomendados en cuantización Q4 sin sacrificar demasiada calidad. El software listado está seleccionado por ser open source, activamente mantenido y compatible con el hardware de este tier.

Guía de configuración paso a paso

  1. 1

    Install whisper.cpp: `git clone https://github.com/ggerganov/whisper.cpp && cd whisper.cpp && make`.

  2. 2

    Download the Large v3 model: `bash ./models/download-ggml-model.sh large-v3`.

  3. 3

    Transcribe an audio file: `./main -m models/ggml-large-v3.bin -f your_audio.mp3`.

  4. 4

    For GPU (CUDA) transcription, compile with: `make WHISPER_CUDA=1`.

  5. 5

    For batch transcription, use the Python mode with the `openai-whisper` library.

Preguntas frecuentes

How accurate is Whisper Large v3?

Whisper Large v3 has a word error rate (WER) comparable to paid cloud services. It reaches very high quality in clear recordings. For medical or legal transcription, always review the output manually.

Do I need a GPU for Whisper?

Not necessarily. With whisper.cpp you can transcribe on CPU with reasonable results. A GPU speeds up processing by 5x–20x. For long files (hours of audio), the GPU makes a significant difference.

Does Whisper support multiple languages?

Yes. Whisper was trained on 99 languages and has excellent multilingual support. Add `--language en` (or the appropriate code) to the command to force the language and improve accuracy.

Found this useful? Get guides like this in your inbox every week.

No spam. Unsubscribe in one click.