OpenAI Whisper transcription online

What is OpenAI Whisper?

OpenAI Whisper is an open-source automatic speech recognition (ASR) model trained on 680,000 hours of multilingual audio collected from the public web. It was released by OpenAI in 2022 and has become a standard reference model for ASR research and production use. Whisper uses a transformer encoder-decoder architecture: the encoder reads an audio spectrogram, and the decoder generates text tokens, allowing it to transcribe or translate audio in a single forward pass.

Unlike earlier ASR systems that required language-specific models, Whisper handles 57 languages with a single model by detecting the language from the first 30 seconds of audio. This makes it practical for multilingual recordings, foreign-language interviews, and international content without any manual configuration. It also supports translation: it can return a transcript directly in English regardless of the source language.

convert.express runs Whisper through the OpenAI API. You upload your file, it is processed server-side, and the transcript is returned to your browser. Files are deleted automatically after 24 hours. The API terms of service prohibit OpenAI from using your audio to train models.

Whisper vs. MAI-Transcribe 1.5 vs. Gemini 3.5 Transcribe

convert.express offers three models. Here is how they compare. Full comparison →

All three models' transcripts can also be translated afterward into any supported language using the dashboard's Translate action — the row above only covers Whisper's free, single-pass translation option at upload time.

Whisper's figures, MAI-Transcribe 1.5's AA-WER figure, and Gemini 3.5 Transcribe's vendor-published figure come from three different benchmarks measured under different conditions, so the word error rate row is not a strict apples-to-apples comparison.

Frequently asked questions

What is OpenAI Whisper and how does it work?

OpenAI Whisper is an open-source automatic speech recognition (ASR) model trained on 680,000 hours of multilingual audio. It converts spoken audio to text by processing the audio spectrogram through a transformer encoder-decoder architecture. convert.express runs Whisper via the OpenAI API — you upload the file, it is processed server-side, and the transcript is returned to you.

How accurate is Whisper transcription?

On clean benchmark English (e.g. LibriSpeech test-clean), published third-party benchmarks put Whisper at roughly 2.7% word error rate (WER) — about 97 words out of every 100 correct. Real-world recordings with background noise, accents, or multiple speakers push that to roughly 8% WER. These are published benchmark figures, not our own measurements, and vary by test set. Technical terminology and unusual proper nouns are the most common sources of error.

Which languages does Whisper support?

Whisper supports 57 languages including Arabic, Chinese, Dutch, English, French, German, Hindi, Indonesian, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish, Turkish, Ukrainian, and more. It includes coverage for minority and regional languages not covered by some commercial ASR systems. The language is detected automatically — you do not need to specify it manually.

Can Whisper translate recordings into English?

Yes. When you select the translation option in convert.express, Whisper transcribes the audio and returns the result in English regardless of the source language, in a single pass at no extra cost. If you need a different target language, or want to translate a transcript you already generated, use the Translate action in your dashboard instead — it works on any completed job, on any of the three models, and supports every language convert.express offers, not just English.

How long does Whisper transcription take?

Whisper processes audio at roughly 60–90 seconds per hour of recording. A 30-minute interview takes around 45–60 seconds to return. A 5-minute voice memo is typically ready in under 15 seconds. Processing time depends on file size and server load.

When should I choose Whisper over MAI-Transcribe 1.5 or Gemini 3.5 Transcribe?

Choose Whisper when: you want a free, single-pass English translation at upload time without an extra step afterward — neither MAI-Transcribe 1.5 nor Gemini 3.5 Transcribe offers this; or you prefer a generally-available model with no preview caveats. MAI-Transcribe 1.5 is faster and offers entity biasing on its 43 supported languages. Gemini 3.5 Transcribe has the widest language coverage, in-model speaker diarization for up to 8 speakers, and up to 1,000 entity-biasing terms, but is preview and cannot translate. All three models' transcripts can be translated into English (or any other supported language) after the fact via the dashboard's Translate action.