← Back to all models

Whisper

OpenAIAudio & SpeechOpen Source

Open-source speech recognition model trained on 680,000 hours of multilingual audio. Capable of transcription and translation across 99 languages.

Abilities

Audio TranscriptionTranslationMultilingualOn-device

Use Cases

EnterpriseAutomationEducation

Available in Tools

Availability

How to Use

Install via pip (`pip install openai-whisper`) and run locally, or use the OpenAI Audio API for cloud-based transcription.

Pros

  • Fully open source with MIT license — complete freedom to use, modify, and distribute
  • Supports 99 languages with automatic language detection
  • Multiple model sizes (tiny 39M to large 1.5B) — trade accuracy for speed as needed
  • Can run fully offline on local hardware — no data leaves your machine
  • Whisper.cpp port enables real-time transcription on CPU-only devices
  • Free to self-host — zero marginal cost per minute of audio processed

Cons

  • Large-v3 model requires ~10GB VRAM — not practical on low-end hardware
  • Real-time transcription needs significant optimization (batching, VAD, streaming)
  • Struggles with heavy accents, fast speech, and domain-specific jargon
  • No built-in speaker diarization — cannot distinguish between different speakers
  • Can hallucinate repetitive text on silent or noisy audio segments

What to Use It For

star

Perfect For

Batch audio transcription in multiple languages

Supports 99 languages with automatic detection — process multilingual audio without manual configuration

Offline speech-to-text on your own hardware

MIT license and local deployment mean no data leaves your machine — ideal for sensitive audio

Transcribing meetings and podcasts

Handles long-form conversational audio well with good punctuation and formatting

thumb_up

Good For

Subtitle generation

Built-in timestamp support produces timed segments ready for SRT/VTT subtitle formats

Voice command recognition

Smaller model variants (tiny, base) provide fast enough inference for near-real-time command detection

warning

Not Recommended

Real-time live captioning without optimization

Out-of-the-box latency is too high — needs VAD, streaming, and batching pipelines to work live

Try instead: cloud ASR services

Speaker identification

No built-in diarization — cannot tell who is speaking, only what is being said

Try instead: pyannote + Whisper

block

Do Not Use For

Text generation

It is an automatic speech recognition model only — it converts audio to text, not the reverse

Try instead: GPT-4o

Music transcription

Trained on speech data — cannot parse musical notes, chords, or instrument separation

Try instead: specialized music AI

Technical Details

PricingFree (self-hosted) / $0.006/min via API
Parameters1.5B (large-v3)
detail.contextWindow30 seconds audio chunks
LicenseMIT

Links