What is Whisper (openai/whisper)?
openai/whisper is the original code release for Whisper, OpenAI's speech recognition model family: a Python package (openai-whisper on PyPI) and a whisper command-line tool that transcribe speech, translate it into English and identify the spoken language. It downloads and runs six sizes, from tiny (39M parameters, about 1 GB of VRAM) to large (1,550M, about 10 GB) and turbo (809M, about 6 GB), plus English-only versions of the four smallest.
This review covers the software. The turbo checkpoint it uses by default is reviewed separately as Whisper large-v3-turbo under models.
The code is MIT-licensed, has about 112,700 GitHub stars and drew about 6.2 million PyPI downloads in the last month. The last PyPI release, v20250625, dates from June 2025; the commits since, the latest on 31 August 2026, are small fixes and a JSONL output format that pip users do not yet get.
Is Whisper (openai/whisper) worth it?
openai/whisper remains the reference every other Whisper runtime is measured against, and it is the slowest of them. In faster-whisper's published benchmark, 13 minutes of audio on large-v2 took 2m23s with openai/whisper on an RTX 3070 Ti, against 1m03s for faster-whisper and 1m05s for whisper.cpp; on a CPU with the small model it took 6m58s against 2m05s for whisper.cpp. Those runs used v20240930, and the code has changed little since.
What it offers is completeness and simplicity. One pip install plus ffmpeg gives a CLI that writes text, subtitles, TSV and JSON, translation into English with the multilingual models, language detection and word-level timestamps. The README flags a trap that catches people: turbo, the default model, is not trained for translation and returns the original language even with --task translate, so translation needs medium or large.
Maintenance is light. Five commits have landed since July 2025, the most useful a fix for beam search falling back to a slow attention kernel and a JSONL output format, both in July 2026; none has reached PyPI. OpenAI has released no new checkpoint since turbo in 2024 and is retiring its hosted whisper-1 API model in February 2027, which does not affect this open code but shows where the company's attention has gone.
The package earns its place as a baseline, for teaching, and for small jobs where a few extra minutes cost nothing. For a transcription service, a back-catalogue batch or a laptop app, the same weights run faster and in less memory elsewhere, and faster-whisper reports the same accuracy.
Where it earns its place
A newsroom developer handling a few interviews a week
Turn recorded interviews into searchable text and subtitles
Run whisper interview.mp3 --model turbo on a desktop with a 6 GB GPU and get text, SRT and JSON files in one pass. There is no per-minute fee, and at this volume the slower speed does not matter, though quotes still need checking against the audio.
A machine-learning engineer evaluating a faster Whisper runtime
Confirm a production engine gives the same transcripts as the original
Transcribe a fixed test set with openai/whisper and the candidate runtime using the same model size and beam settings, then compare word error rates. The reference code is small and readable, which makes discrepancies easy to trace.
A university lab transcribing Japanese and Spanish field recordings
Get English translations alongside the original-language transcripts
Use the medium or large model with --task translate, since turbo ignores translation. Expect about 10 GB of VRAM for large and a slower run; accuracy varies widely by language, as OpenAI's own per-language chart shows.
Before you commit
- Setup
- pip install -U openai-whisper plus ffmpeg from your package manager; Rust may be needed if tiktoken has no prebuilt wheel for your platform. Models download on first use, and a first transcript takes minutes.
- Integrations
- The whisper CLI and a Python API (load_model, transcribe, detect_language); outputs txt, vtt, srt, tsv and json, with jsonl on the main branch. Runs on PyTorch with CUDA or CPU.
- Data and privacy
- Runs locally; audio never leaves the machine once the weights are downloaded. The MIT licence covers both code and weights.
- Cost at scale
- Free; the cost is GPU time, which it uses less efficiently than faster-whisper, 2m23s against 1m03s for the same 13-minute file in faster-whisper's benchmark.
- Limits
- No PyPI release since June 2025; turbo cannot translate; the slowest of the common runtimes; accuracy varies sharply by language.
What’s good
- One pip install gives a full CLI and Python API
- All six model sizes, translation and language detection
- MIT licence; about 6.2 million PyPI downloads a month
- The reference other Whisper runtimes are checked against
What’s not
- More than twice as slow as faster-whisper on GPU in published benchmarks
- No PyPI release since June 2025
- The default turbo model cannot translate
- Needs ffmpeg, and sometimes Rust, to install
Best for
Developers and researchers who want the reference Whisper implementation for small jobs, teaching or checking another runtime's output.
Whisper (openai/whisper) alternatives
Whisper (openai/whisper): common questions
- Is Whisper (openai/whisper) worth it?
- The Whisper package is still the shortest path from an audio file to a transcript, but for any real volume faster-whisper or whisper.cpp do the same job in a fraction of the time. We rate it 3.6 out of 5.
- How much does Whisper (openai/whisper) cost?
- Free and open source under the MIT licence, code and weights alike. No paid edition.
- What are the downsides of Whisper (openai/whisper)?
- More than twice as slow as faster-whisper on GPU in published benchmarks; no PyPI release since June 2025; the default turbo model cannot translate; needs ffmpeg, and sometimes Rust, to install.
- Who is Whisper (openai/whisper) best for?
- Developers and researchers who want the reference Whisper implementation for small jobs, teaching or checking another runtime's output.
- How much does Whisper (openai/whisper) cost for a team or in production?
- Free; the cost is GPU time, which it uses less efficiently than faster-whisper, 2m23s against 1m03s for the same 13-minute file in faster-whisper's benchmark.
- Is Whisper (openai/whisper) safe for company data?
- Runs locally; audio never leaves the machine once the weights are downloaded. The MIT licence covers both code and weights.
- What are the best Whisper (openai/whisper) alternatives?
- The closest alternatives we have reviewed are audio.cpp and VibeVoice.
Answers checked Nov 14, 2026.