OpenAI Whisper

by OpenAI

OpenAI Whisper: The reference open-source multilingual ASR model from OpenAI. Skip the install: WhipScribe transcribes it in your browser. Packs: $4 for 500 minutes, $12 for 2,000, $24 for 5,000; Workspace $19 a month. Up to 10 hours or 5 GB per file. Measured turnaround: about 2 minutes per audio hour. 99 languages. Checked 25 Oct 2026.

Don’t want to install OpenAI Whisper? Drop the file here.

Every recording starts with a free preview of its opening minutes. Credits never expire.

The reference Whisper from OpenAI — weak-supervised multilingual ASR, MIT-licensed, 99 languages. Six model sizes plus the new large-v3-turbo. Pick a size, copy a recipe.

Linux, macOS, Windows, GPU  ·  free

The pip package name is openai-whisper, not whisper

pip install whisper installs an unrelated project, Graphite's round-robin database. OpenAI's release on PyPI is openai-whisper, latest 20250625 (uploaded 26 Jun 2025), for Python 3.8 or newer. In Python you still import whisper.

If pip fails while building tiktoken, the README says to install Rust, or run pip install setuptools-rust. To install the newest commit instead of the release: pip install git+https://github.com/openai/whisper.git

  1. Install the package and FFmpeg (macOS with Homebrew, Ubuntu or Debian, Windows with Chocolatey).

    pip install -U openai-whisper
brew install ffmpeg
sudo apt update && sudo apt install ffmpeg
choco install ffmpeg

  2. Transcribe from the command line. turbo is the default model.

    whisper audio.mp3 --model turbo
whisper japanese.wav --language Japanese

  3. Or from Python.

    import whisper

model = whisper.load_model("turbo")
result = model.transcribe("audio.mp3")
print(result["text"])

Sources:

  1. PyPI · openai-whisper (20250625), fetched 26 Oct 2026
  2. PyPI · whisper (Graphite's round-robin database), fetched 26 Oct 2026
  3. openai/whisper · README (GitHub), fetched 26 Oct 2026

Whisper models: size, VRAM, speed and direct download links

From the README, measured by OpenAI on English speech with an A100; speed is relative to large:

turbo is an optimised large-v3 and the command-line default, but it is not trained for translation: with --task translate it returns the original language. Use medium or large to translate into English.

whisper.load_model() downloads checkpoints to ~/.cache/whisper. The direct links, from whisper/__init__.py:

SizeParametersEnglish-onlyMultilingualVRAMRelative speed
tiny39 Mtiny.entiny~1 GB~10x
base74 Mbase.enbase~1 GB~7x
small244 Msmall.ensmall~2 GB~4x
medium769 Mmedium.enmedium~5 GB~2x
large1550 MN/Alarge~10 GB1x
turbo809 MN/Aturbo~6 GB~8x

Sources:

  1. openai/whisper · README (GitHub), fetched 26 Oct 2026
  2. whisper/__init__.py · checkpoint URLs and cache folder, fetched 26 Oct 2026

OpenAI Whisper .pt checkpoint URLs

The long folder name in each URL is the file's SHA-256, which load_model checks after the download.

ModelDirect download
tiny.entiny.en.pt
tinytiny.pt
basebase.pt
smallsmall.pt
mediummedium.pt
large-v3large-v3.pt
large-v3-turbo (turbo)large-v3-turbo.pt

Sources:

  1. whisper/__init__.py · checkpoint URLs and cache folder, fetched 26 Oct 2026

Which languages OpenAI Whisper supports

The tokenizer lists 100 language codes (whisper/tokenizer.py, read 26 Oct 2026). Pass one with --language, or let Whisper detect it from the first 30 seconds. Accuracy varies widely by language; the README charts word and character error rates for large-v2 and large-v3 on Common Voice 15 and Fleurs.

Sources:

  1. whisper/tokenizer.py · LANGUAGES, fetched 26 Oct 2026
  2. openai/whisper · README (GitHub), fetched 26 Oct 2026
TL;DR

OpenAI's openai-whisper Python package is the canonical reference implementation of Whisper — a single large encoder-decoder transformer in PyTorch, weak-supervised on 680,000 hours of multilingual audio (paper: Radford et al., 2022). Ships six base checkpoints — tiny (39M), base (74M), small (244M), medium (769M), large (1.55B), and the newer turbo (809M, ~8× faster than large) — and supports 99 languages for transcription plus translation-to-English.

Best for research baselines, anyone learning the ASR space, and quick local experiments where dependency surface matters more than throughput. For production speed pick a derivative — faster-whisper (CTranslate2, ~4× faster) or whisper.cpp (C/C++, fastest on Apple Silicon and CPU). License: MIT on both code and weights. Newest checkpoint is large-v3-turbo — 4 decoder layers instead of 32, near-large-v3 quality at a fraction of the wall-clock.

Category
Open source
License
MIT
Stars
★ 111.4k
Last push
2026-08-31
Pricing
free
Platforms
Linux, macOS, Windows, GPU

What it is

Whisper is OpenAI's flagship speech-to-text model, trained on 680,000 hours of multilingual audio. The original Python package is the reference implementation — easy to install, but intentionally simple. For production speed you almost always want a derivative (whisper.cpp, faster-whisper, whisperX) that wraps the same weights in a faster runtime. Free, MIT-licensed, runs on CPU or GPU.

Best for: Research, baseline accuracy, teams that want the canonical reference implementation.
Watch out for: Pure-Python inference is slow on CPU; no built-in speaker diarization; batch-only, no streaming.

Install / use

pip install -U openai-whisper

Setting OpenAI Whisper up or fixing an error? OpenAI Whisper setup guide and common fixes.

Pick a model · 6 checkpoints

openai-whisper ships six base model sizes. Bigger = more accurate, more VRAM, slower. The table below mirrors the canonical README. Each card links the GitHub README anchor or the official Hugging Face model card — never a direct weights download. The Python package fetches weights on first use.

tiny
Smallest · ~10× speed · edge / smoke test

39M parameters. The fastest checkpoint and the easiest to run on a CPU laptop or edge box. WER trails the larger models substantially — fine for smoke tests, language ID, or short-clip preview. A separate tiny.en English-only variant ships alongside.

tiny
39M params · ~1 GB VRAM · ~10× relative speed vs large · multilingual + tiny.en variant
base
~7× speed · usable CPU default

74M parameters. The lowest checkpoint that produces broadly usable English transcripts on clean audio. Still fits inside ~1 GB VRAM and runs comfortably on consumer CPUs. A base.en English-only variant is also published.

base
74M params · ~1 GB VRAM · ~7× relative speed vs large · multilingual + base.en variant
small
~4× speed · CPU sweet spot

244M parameters. The point at which English WER on clean audio gets close to medium for most use cases. Comfortable on a 2 GB GPU or a modern CPU. A small.en English-only variant is published alongside.

small
244M params · ~2 GB VRAM · ~4× relative speed vs large · multilingual + small.en variant
medium
~2× speed · accuracy/VRAM balance

769M parameters. Strong multilingual accuracy on a 6–8 GB GPU. Common default for desktop transcription where large won't fit alongside the rest of the stack. A medium.en English-only variant is published alongside.

medium
769M params · ~5 GB VRAM · ~2× relative speed vs large · multilingual + medium.en variant
large-v3
Flagship accuracy · 99 languages

1.55B parameters. The full large encoder-decoder — best WER in the family and the baseline most papers compare against. Needs ~10 GB VRAM at fp16. Reference card: openai/whisper-large-v3 on Hugging Face.

large-v3
1.55B params · ~10 GB VRAM · 1× relative speed (baseline) · 99 languages · multilingual only
large-v3-turbo
Newer · ~8× speed at near-large quality

809M parameters. A 2024 release derived from large-v3 by collapsing the decoder from 32 layers to 4. Roughly 8× faster than large at near-equivalent transcription quality. Note: not trained for translation. Reference card: openai/whisper-large-v3-turbo on Hugging Face.

turbo
809M params · ~6 GB VRAM · ~8× relative speed · 99 languages · transcription only (no translate task)
Selecting at runtime: pass the model name as the first arg to whisper.load_model("turbo") or the --model CLI flag. Weights download from OpenAI's CDN on first use and cache under ~/.cache/whisper. For production throughput on the same weights see faster-whisper (CTranslate2) or whisper.cpp (C/C++).

Setup recipes · pick one and copy

Three runnable configurations from the openai/whisper README. Each block runs against the latest tagged release (v20250625 or later) and Python ≥3.9.

1Install + first transcript (CLI)

pip install · ffmpeg required · single-command transcription with SRT output. The default starting point.

# openai-whisper (latest, v20250625+) · Python >=3.9
# 1) ffmpeg is a hard runtime dep — install it first.
#    macOS:    brew install ffmpeg
#    Ubuntu:   sudo apt update && sudo apt install ffmpeg
#    Windows:  choco install ffmpeg     (or scoop install ffmpeg)

pip install -U openai-whisper

# 2) Transcribe a file directly to SRT subtitles.
whisper input.mp3 --model turbo --output_format srt

# Other useful flags:
#   --language Japanese        force a language (skip auto-detect)
#   --task translate           translate to English (not supported by 'turbo')
#   --device cuda              run on GPU
#   --output_format all        write txt + srt + vtt + tsv + json
Source: openai/whisper · Command-line usage ↗. PyPI: openai-whisper ↗
2Python API + word timestamps

load_model() · transcribe() with word_timestamps=True · iterate segments and per-word timings.

# Python >=3.9, openai-whisper installed.
import whisper

model = whisper.load_model("turbo")   # or tiny / base / small / medium / large-v3

result = model.transcribe(
    "input.mp3",
    word_timestamps=True,
    # language="en",                 # skip auto-detect if you know it
    # task="translate",              # translate to English (not on 'turbo')
)

print(result["text"])
for seg in result["segments"]:
    print(f"[{seg['start']:6.2f} -> {seg['end']:6.2f}] {seg['text']}")
    for w in seg.get("words", []):
        print(f"   {w['start']:6.2f}-{w['end']:6.2f}  {w['word']!r}")
Source: openai/whisper · Python usage ↗. Word-timestamp accuracy drifts on long-form audio — for tighter per-word timing see WhisperX (forced alignment).
3GPU + half-precision

device="cuda" + fp16=True · the production knobs on the reference implementation. Cuts VRAM and wall-clock roughly in half on supported GPUs.

# CUDA toolkit + a matching torch build already installed.
import whisper

model = whisper.load_model("large-v3", device="cuda")

result = model.transcribe(
    "podcast.wav",
    fp16=True,                 # default True on CUDA; explicit here for clarity
    beam_size=5,
    best_of=5,
    temperature=0.0,
    # condition_on_previous_text=False,  # try if you see hallucination loops
)

print(result["text"][:500])
Reference: transcribe.py options ↗. For ~4× faster GPU inference on the same weights, switch to faster-whisper.

What people actually do with OpenAI Whisper-style transcription

The tool is the means. These are the jobs — each one priced at published rates, each one wired up on its own page.

Transcribe a whole back catalogue in one runThe archive is the asset, but one-at-a-time upload means it never gets done.…1,000 hours ≈ $480, onceTranscribe every new file in a Drive folderFiles accumulate in a folder nobody opens, and the knowledge in them is invisible to search.…≈ $0.29 per audio hour on the $24 packCut short-form clips out of a webinar automaticallyThe clips worth posting are buried somewhere in a 60-minute recording and finding them by scrubbing costs more than the…60-min webinar ≈ $0.48

Wire it into Amazon S3n8nMake.comGoogle DriveZapierSlack

Features

Speaker diarizationNo
Word-level timestampsYes
Streaming / real-timeNo
Languages supported99
HIPAA eligibleNo

Links

OpenAI Whisper vs Whipscribe

FeatureOpenAI WhisperWhipscribe
CategoryOpen sourceTranscription APIs
Pricingfree$4–$24 one-time packs (credits never expire) · free preview of the opening minutes
Speaker identificationNoNamed in the summary, key quotes and chapters
Word timestampsYesYes
StreamingNoNo
Languages9999
PlatformsLinux, macOS, Windows, GPUWeb, API, MCP

Alternatives to OpenAI Whisper

whisper.cpp
Georgi Gerganov

C/C++ port of Whisper — runs on anything, from a Raspberry Pi to Apple Silicon.

OSS · MIT ★ 54.7k
faster-whisper
SYSTRAN

4× faster than reference Whisper using CTranslate2 — production sweet spot.

OSS · MIT ★ 26.2k
whisperX
Max Bain

Faster-whisper + forced alignment + speaker diarization in one pipeline.

OSS · BSD‑2‑Clause ★ 23.9k

Frequently asked about OpenAI Whisper

What is the pip package name for OpenAI Whisper?

openai-whisper. Run pip install -U openai-whisper, then import whisper in Python. The package called whisper on PyPI is a different project, Graphite's round-robin database.

Which languages does OpenAI Whisper support?

100 language codes in whisper/tokenizer.py. Set one with --language, or let Whisper detect it.

Where can I download the Whisper .pt model files directly?

The URLs are in whisper/__init__.py, on openaipublic.azureedge.net; the table on this page lists them. whisper.load_model() fetches them to ~/.cache/whisper for you.

Is OpenAI Whisper free?

Yes. The openai-whisper package and all released model weights are MIT-licensed, free for commercial and non-commercial use. Inference cost is whatever hardware you run it on.

Does Whisper support speaker diarization?

No. Vanilla Whisper outputs text + segment timestamps but does not label speakers. To get 'who said what,' pair it with a diarization library (e.g. pyannote) or use whisperX, which bundles both.

What's the difference between Whisper and faster-whisper?

Same underlying model weights; different runtime. faster-whisper uses CTranslate2 and is roughly 4x faster on GPU with lower VRAM use. Accuracy is essentially identical. For production, faster-whisper is usually the better choice.

Can Whisper run on CPU?

Yes, but it's slow. Real-time factor on a modern laptop CPU with the large-v3 model is well below 1x. For CPU-bound workloads, whisper.cpp is dramatically faster than the reference Python implementation.

Does Whisper produce word-level timestamps?

The reference implementation has a word_timestamps flag, but timings can drift on long-form audio. For more accurate per-word timing, use whisperX (forced alignment) or stable-ts.

Does OpenAI Whisper work on Mac, Windows and Linux?

OpenAI Whisper runs on Linux, macOS, Windows and GPU.

Can OpenAI Whisper transcribe live audio?

No — OpenAI Whisper works on finished files, not a live stream.

Is OpenAI Whisper HIPAA compliant?

OpenAI Whisper does not claim HIPAA compliance. For PHI, run it on infrastructure you control or choose a service that will sign a BAA.

What are the limitations of OpenAI Whisper?

Pure-Python inference is slow on CPU; no built-in speaker diarization; batch-only, no streaming.

Who is OpenAI Whisper best for?

Research, baseline accuracy, teams that want the canonical reference implementation.

What should I know before choosing OpenAI Whisper?

Best for research baselines, anyone learning the ASR space, and quick local experiments where dependency surface matters more than throughput. For production speed pick a derivative — faster-whisper (CTranslate2, ~4× faster) or whisper.cpp (C/C++, fastest on Apple Silicon and CPU). License: MIT on both code and weights. Newest checkpoint is large-v3-turbo — 4 decoder layers instead of 32, near-large-v3 quality at a fraction of the wall-clock.

Does OpenAI Whisper run offline?

OpenAI Whisper runs on your own machine — it is open source, so nothing leaves the computer unless you configure it to.

Whipscribe is a hosted transcription service. If you want transcripts without running infrastructure, paste a URL or drop a file in the form below — you'll have a transcript in seconds.

What we measured, last 30 days

2,547 recordings, 1,244 audio hours, 66 languages detected. An hour of audio came back in 3.5 minutes for half of the recordings and within 18.0 minutes for nine in ten, measured from the moment the file landed to the finished transcript. Refreshed 2026-10-07; method on how we test.

Transcribe your recordings

Running OpenAI Whisper yourself? Send the same file here and compare the transcript. Transcribe any language, any recording · Transcription services from $0.29 an hour · Transcribe from any meeting or recording platform · How to transcribe, step by step

Explore

All transcription tools · By language · By platform · By file format · How-to guides · Transcription services · Transcription API · Pricing

Related

Have the recording already?

Skip the setup and send the file or the link: see how to transcribe an MP3 of any length, transcribing a podcast episode and the SRT file generator.

Rather have it done for you?

Free tools cost setup time. When you want a finished transcript instead: see transcription services from $4.