# Voice and audio

Turn recordings into accurate text, label who spoke, read text aloud in a natural voice, and split a finished mix back into its parts.

## Models for this work

| Model | Best for | Quality | Speed | From |
|---|---|---|---|---|
| [Whisper large-v3](https://gpuvault.io/model/whisper-large-v3/) | Any job that turns recorded speech into text | Excellent | Very fast | $0.45/hr |
| [pyannote speaker diarization](https://gpuvault.io/model/pyannote-diarization/) | Adding speaker labels to a transcript | Very good | Very fast | $0.22/hr |
| [Kokoro TTS](https://gpuvault.io/model/kokoro-82m/) | Narration and voiceover at volume, for almost nothing | Very good | Very fast | $0.22/hr |
| [Demucs](https://gpuvault.io/model/demucs/) | Pulling one sound out of a mix that was never separate | Excellent | Fast | $0.45/hr |

## What do you need to do?

- **Transcribe audio** — Recordings into accurate, timestamped text. https://gpuvault.io/model/whisper-large-v3/
- **Identify speakers** — Who said what, through a whole meeting. https://gpuvault.io/model/pyannote-diarization/
- **Generate narration** — Text read aloud in a natural voice. https://gpuvault.io/model/kokoro-82m/
- **Subtitles** — Timed caption files for an edit. https://gpuvault.io/model/whisper-large-v3/
- **Translate speech** — Foreign language audio into English text. https://gpuvault.io/model/whisper-large-v3/
- **Separate a mix** — Vocals, drums, and dialogue pulled apart. https://gpuvault.io/model/demucs/

## Questions people ask

### How accurate is the transcription?

On clear recorded speech in English, Whisper large-v3 typically gets 95% to 98% of words right, which is close to a professional human transcriber. Accuracy drops on heavy background noise, strong crosstalk, and specialist vocabulary such as drug names or company jargon. It handles accents well. Expect to proofread technical terms and proper nouns, not the general text.

### How much does it cost to transcribe 40 hours of audio?

About $0.60. Whisper transcribes an hour of audio in roughly 2 minutes on an RTX 4090 at $0.45 per hour, so 40 hours takes around 80 minutes of GPU time. Commercial transcription services charge $0.25 to $1.50 per minute of audio, which puts the same job between $600 and $3,600.

### Will it tell me who is speaking?

Yes, if you ask for it. Whisper writes down the words but does not separate speakers on its own. Tick "identify speakers" on the transcription preset and we run pyannote alongside it, which labels each segment as Speaker 1, Speaker 2 and so on with timestamps. You attach real names afterwards. It costs a few cents more and roughly a minute per hour of audio.

### What languages does it handle?

Around a hundred, and it works out which one it is hearing without being told. Accuracy is highest in English, Spanish, Italian, German, Japanese, and Portuguese, and falls off for languages with less recorded material available. It will also translate as it transcribes, so a French interview can come back as English text in a single pass at no extra cost.

### Can I use the generated voices commercially?

Yes. Kokoro is Apache 2.0 licensed, and there are no restrictions on selling or publishing the audio it produces. Whisper and Demucs are likewise permissive. Note that Kokoro is not a voice cloning tool and will not reproduce a specific person, which is deliberate.

### What file formats can I upload?

Any common audio or video file. MP3, WAV, M4A, FLAC, MP4, MOV, and MKV all work, and we pull the audio out of video files for you. There is no need to convert anything first. For the best accuracy, upload the original recording rather than a heavily compressed copy, since compression removes exactly the detail the model uses.

---

*Source: https://gpuvault.io/working-on/voice-and-audio/ — GPUVault, GPU rental by the minute.*