Whisper

Name: Whisper
Author: OpenAI

Wizper

byOpenAI

Transcribe multilingual speech with word-level precision or translate spoken audio directly into clean, punctuated English text

Transcribe audio

How Whisper works

Convert raw audio files into clean, punctuated transcripts or English translations in three simple steps.

Upload audio recording

Provide a clear mono or stereo audio file under 25 MB in MP3, AAC, or WAV format.

Select task and language

Choose whether to transcribe the original language or translate to English, with auto-detection across nearly 100 languages.

Receive punctuated transcript

Get clean, publication-ready text with standardized capitalization, punctuation, and normalized speech fillers.

What Whisper is good at

Engineered on an optimized 1.55-billion parameter architecture for high-speed speech recognition across nearly 100 languages.

Multilingual Speech Recognition

Transcribe spoken audio across nearly 100 languages with automatic language detection, maintaining high accuracy across varied accents and noisy background environments.

Direct-to-English Translation

Convert foreign-language spoken dialogue directly into fluent, punctuated English text in a single pass without chaining separate translation pipelines.

Accelerated Inference Performance

Process long recordings at more than double the speed of baseline models while preserving identical word error rates on 1.55-billion parameter weights.

Automated Formatting and Punctuation

Transform continuous speech into structured paragraphs complete with grammatical punctuation, proper noun capitalization, and stripped filler disfluencies.

Made with Whisper

Explore real-world transcription and translation outputs across documentary audio, broadcast media, multi-speaker interviews, and video captions.

Archival interview restoration and documentary transcription

Direct-to-English brand campaign translation

Social media video captioning under heavy background noise

Technical keynote address and symposium transcription

Cinematic film dialogue transcription and subtitling

What people build with Whisper

Built for video creators, documentary filmmakers, podcasters, and localization teams who need fast, verbatim-accurate speech-to-text.

Video Subtitling and Social Clips

Generate accurate, time-efficient captions for short-form video creators and post-production editors, turning spoken social content into clean, readable on-screen text.

Documentary and Archival Research

Transcribe historical recordings, field interviews, and oral history archives across diverse global accents and varying recording conditions.

Global Advertising Localization

Translate regional voiceover tracks and international consumer testimonials directly to English text to accelerate multinational campaign reviews.

Film Dialogue and Indie Post-Production

Extract accurate dialogue transcripts from production scratch tracks and location audio to create master subtitle files and edit decision lists.

Podcast and Editorial Publishing

Turn multi-hour spoken audio interviews into clean, readable text ready for editorial indexing, show notes, and article drafting.

Whisper specs

Plans and credits

Inputs: Audio
Output: text
Cost: 35 credits per run

Frequently asked questions

Wizper is an ultra-optimized serverless implementation of OpenAI's Whisper v3 Large model. It delivers the exact same Word Error Rate (WER) and transcription accuracy as stock Whisper v3 Large (1.55 billion parameters) while running at more than double the inference speed at lower computational cost.

Audio files should be under 25 MB in standard formats such as MP3, AAC, or WAV. For optimal transcription accuracy, use clear mono recordings sampled at 16 kHz or higher, and pre-filter prolonged stretches of dead silence to prevent repetition loops.

No, the translation task is strictly unidirectional from foreign languages into English. The model can transcribe spoken speech in nearly 100 native languages, but its translation mode is designed solely to translate non-English speech directly into English text.

No, the model does not natively perform speaker diarization or label distinct speakers. It outputs all recognized speech into a continuous, punctuated transcript, so multi-speaker dialogue with heavy overlap may drop words if not separated beforehand.

Leaving the language setting on Auto-detect enables automatic language identification across nearly 100 languages. For audio clips shorter than 5 seconds or recordings with phonetically similar dialects, explicitly passing the language code eliminates detection latency and prevents misclassification.