Transcribe 100 hours of audio with Whisper for USD 0.45
Example project: turn 100 hours of calls or episodes into searchable text with Whisper large-v3 on a rented GPU, priced at today's rates.
- Whisper large-v3 transcribes 100 hours of audio in about 1.5 hours of GPU time: USD 0.45 today, about USD 0.01 per hour of audio.
- It handles about a hundred languages, Hindi, Tamil, Bengali, Japanese and Korean among them.
- You pay for GPU time, not per minute of audio, and nothing once the machine is gone.
The challenge
A call centre, a podcast network or a newsroom keeps hundreds of hours of recordings that nobody can search. Paid transcription services charge per minute, and audio is often in Hindi, Tamil, Japanese or several languages at once.
Why a rented GPU
- Transcription is a batch job: rent a GPU for an hour or two, then stop paying.
- Whisper large-v3 is open and handles about a hundred languages, so one setup covers mixed-language audio.
How it works, step by step
- Upload the recordings to a rented GPU running the Whisper speech-to-text software.
- Transcribe them in batches; the text comes back with timestamps.
- Spot-check a few transcripts, especially names and numbers.
- Download the transcripts and load them into your search tool or help desk, then destroy the machine.
Which GPU to choose
The cheapest way today is RTX 3090: 1.5 hours for USD 0.45.
The quickest, H100 SXM, takes 0.6 hours for USD 2.24.
Upload time matters with many gigabytes of audio: pick a host with a fast download speed, shown on every offer.
What it costs today
| Step | GPU | Time | Cost |
|---|---|---|---|
| Transcribe 100 hours of audio (Whisper large-v3) | RTX 3090 | 1.5 h | USD 0.45 |
| The same, on the quickest card | H100 SXM | 0.6 h | USD 2.24 |
What you end up with
- About USD 0.01 per hour of audio, instead of a per-minute fee.
- Text you can search, quote and translate, with timestamps back to the recording.
- Nothing to install and nothing left running afterwards.
Common mistakes to avoid
- Uploading audio one file at a time from a slow connection: zip it, or pick a host near you.
- Not telling Whisper the language of single-language audio: it guesses, and short clips can come out in the wrong language.
- Forgetting to download the transcripts before destroying the machine.
Keep the bill where you want it
- Billing by the second; a stopped machine pays only for its disk.
- Daily and monthly spending limits, a timer on each machine and an idle stop.
- Prices and payment in your currency, by card.
Questions
How long does it take to transcribe 1 hour of audio?
With Whisper large-v3 on RTX 3090, about 1 minute of GPU time per hour of audio, plus the upload.
Which languages work?
Whisper large-v3 transcribes about a hundred languages, Hindi, Bengali, Tamil, Japanese and Korean among them. Accuracy depends on the language and the audio; try a few recordings first.
How much does it cost to transcribe 100 hours of audio?
About USD 0.45 of GPU time on PowerPod today.