How much does Whisper cost to use?

Whisper is free to self-host with open weights. If you use the OpenAI API, the cost is $0.006 per minute. For those with engineering capacity, self-hosting results in effectively zero per-hour cost.

Does Whisper support multiple languages?

Yes, Whisper is multilingual and supports 99 languages. It is considered the de-facto baseline for speech recognition and is genuinely competitive with paid alternatives on accuracy, particularly the large-v3 release.

Can Whisper identify different speakers automatically?

No, the model has no built-in diarisation. Speaker labels require a separate pipeline, such as pyannote. If you need turnkey diarisation, summarisation, or streaming, AssemblyAI is a better alternative.

Is Whisper easy to integrate into my project?

Integration is straightforward in Python using Hugging Face Transformers. It is available via OpenAI's API for those who do not want to operate a GPU, making it accessible for various engineering capacities.

Are there any known issues with Whisper?

Hallucinations on silent segments are a known issue that requires post-processing to clean up. While these are solvable in production pipelines, they can be surprising out of the box.

Whisper

✓ Editorially verified

OpenAI's open-source speech-to-text — the de-facto baseline.

Free· Free open weights; $0.006/min via OpenAI APIAudioWhisper large-v38.6 / 10

Visit website →

In short

Whisper is OpenAI's free, open-source speech-to-text model supporting 99 languages. It offers strong baseline accuracy for self-hosting or via API, though it lacks built-in diarisation.

Best for

Pick Whisper when you can self-host (or the OpenAI API is fine) and want strong baseline transcription at near-zero per-hour cost.

Skip if

Skip it when you need turnkey diarisation, summarisation, or streaming — AssemblyAI is built for that.

Whisper is OpenAI's open-source speech recognition model. It's free to self-host, multilingual (99 languages), and the baseline against which every other STT model is measured. The large-v3 release is genuinely competitive with paid alternatives on accuracy.

For anyone with engineering capacity, Whisper is the default. Self-hosted, it costs effectively nothing per hour of audio. Available via OpenAI's API for those who don't want to operate a GPU. Hugging Face Transformers makes integration straightforward in Python.

The model has no built-in diarisation — speaker labels need a separate pipeline (pyannote, etc.). Hallucinations on silent segments are a known issue and require post-processing to clean up. For production pipelines these are solvable; out of the box they're surprising.

Editor's take

Whisper is the rare OpenAI release that's open-weight and excellent. It set the standard for what speech-to-text should cost, and it remains the right default for almost any team with engineering capacity.

— The AI Tool Bible editorial team

Pros

✅ Free, open weights
✅ Multilingual (99 languages)
✅ Strong baseline accuracy
✅ Available via API or self-host

Cons

⚠️ No diarisation built in
⚠️ Hallucinations on silent segments

Use cases

transcriptionself-hostedmultilingual

Frequently asked

How much does Whisper cost to use?: Whisper is free to self-host with open weights. If you use the OpenAI API, the cost is $0.006 per minute. For those with engineering capacity, self-hosting results in effectively zero per-hour cost.
Does Whisper support multiple languages?: Yes, Whisper is multilingual and supports 99 languages. It is considered the de-facto baseline for speech recognition and is genuinely competitive with paid alternatives on accuracy, particularly the large-v3 release.
Can Whisper identify different speakers automatically?: No, the model has no built-in diarisation. Speaker labels require a separate pipeline, such as pyannote. If you need turnkey diarisation, summarisation, or streaming, AssemblyAI is a better alternative.
Is Whisper easy to integrate into my project?: Integration is straightforward in Python using Hugging Face Transformers. It is available via OpenAI's API for those who do not want to operate a GPU, making it accessible for various engineering capacities.
Are there any known issues with Whisper?: Hallucinations on silent segments are a known issue that requires post-processing to clean up. While these are solvable in production pipelines, they can be surprising out of the box.