Whisper
github.com
- Free plan
- Open source
- Windows
- macOS
- Linux
- API
- Self-hosted
What Whisper is
Whisper is a free, open-source speech recognition model for multilingual transcription, translation, language identification, and voice activity detection. It can translate speech from multiple languages into English, and documented command-line and Python interfaces support transcribing audio files. Six model sizes—tiny, base, small, medium, large, and turbo—offer different speed and memory tradeoffs; stated VRAM requirements range from approximately 1 GB to 10 GB depending on the model. Whisper runs on Linux, macOS, Windows, and self-hosted deployments. Installation is documented through pip and requires the ffmpeg command-line tool; Rust may be needed if a prebuilt tiktoken wheel is unavailable. The code and model weights are released under the MIT License. OpenAI says the model was trained on 680,000 hours of multilingual, multitask supervised data collected from the web. Performance varies widely by language, and Whisper does not outperform models specialized for LibriSpeech. Turbo is not trained for translation, so multilingual models are recommended for translating non-English speech into English.
Who it suits
Whisper suits developers and researchers building applications or pursuing further work in speech processing. It may also suit users comfortable installing and running models locally through Python and ffmpeg.
Good to know, and what to watch
- Supports transcription, translation, language identification, and voice activity detection.
- Offers six model sizes with different speed and memory tradeoffs.
- Code and weights are released under the MIT License.
- Runs on Linux, macOS, Windows, and self-hosted deployments.
- Installation requires the ffmpeg command-line tool.
- Rust may be needed if a prebuilt tiktoken wheel is unavailable.
- Performance varies widely by language.
- Turbo is not trained for translation tasks.
Plans
Every plan and priceSelf-hosted Whisper
Whisper (open-source code and model weights)
Whisper API
Open source
Getting started
- Open the Whisper GitHub repository
- Install the package through pip
- Install the ffmpeg command-line tool
- Choose a model size from tiny, base, small, medium, large, or turbo
- Use the documented Python or command-line interface to transcribe audio
Limits to check
Model choice determines speed and memory use, with approximately 1 GB to 10 GB of VRAM required. Turbo is not trained for translation tasks; use a multilingual model to translate non-English speech into English.
The Club's review
Overview
Whisper is a locally runnable, open-source speech model for turning multilingual audio into text, translating speech into English, identifying languages, and detecting voice activity. It is aimed at developers and researchers building speech applications or experimenting with robust speech processing. The main decision is whether its broad task coverage and local control outweigh model-specific hardware needs and uneven language performance.
Key features
Whisper supports multilingual transcription, spoken-language identification, speech translation, and voice activity detection. Timestamp support makes its output useful for synchronised text, while TXT, VTT, SRT, TSV, JSON, and JSONL exports cover readable transcripts, subtitle files, and structured processing pipelines.
OpenAI says training used 680,000 hours of multilingual, multitask supervised data collected from the web. That breadth makes Whisper a practical general-purpose foundation, but it is not a guarantee of equal results across languages: performance varies widely, and models specialised for LibriSpeech outperform it on that benchmark. Choose it for coverage and flexibility rather than assuming it is the best option for a particular language or narrow recognition task.
There are six model sizes: tiny, base, small, medium, large, and turbo. They trade speed against memory, with approximately 1 GB to 10 GB of VRAM required depending on the model. Smaller choices are the realistic starting point on constrained hardware; larger choices provide a heavier local workload. Turbo is not trained for translation, so translating non-English speech into English requires one of the multilingual models instead.
The repository documents both command-line and Python interfaces for audio-file transcription. Installation uses pip and requires the ffmpeg command-line tool; Rust may also be needed when a prebuilt tiktoken wheel is unavailable. A Google Colab example provides a hosted starting point, and GitHub Discussions invite third-party extensions and integrations. The code and model weights use the MIT License, allowing the project to serve as a base for applications and further research.
Pricing
Whisper (open-source code and model weights)
free. This plan includes MIT-licensed code and model weights for local installation with Python and ffmpeg. It suits teams that can manage their own environment and hardware and want to avoid a hosted transcription bill. The tradeoff is operational responsibility: installation, compatible dependencies, and the VRAM needed by the selected model remain yours. No paid tier, seat allowance, usage quota, trial period, or renewal term is stated.
Platforms
Whisper runs on Linux, macOS, Windows, and self-hosted deployments. Its command-line and Python interfaces make it a fit for scripted or application-based workflows, while local execution keeps deployment under the user's control. Hardware requirements vary materially by model, from about 1 GB to 10 GB of VRAM.
Who it's for
Whisper is a strong starting point for developers and researchers who need multilingual transcription, English translation, language identification, or voice activity detection and want open code and weights they can install locally. It also suits workflows that need timestamped subtitles or machine-readable exports in several formats.
It is less suitable when a project depends on consistently strong results in one language, targets a LibriSpeech-specialised benchmark, lacks the hardware for its preferred model, or needs translation while using turbo. Those users should consider a language-specialised or hosted service instead.
Pros and cons
- Pro: One MIT-licensed package covers transcription, English speech translation, language identification, and voice activity detection.
- Pro: Six model sizes let users trade speed and VRAM use, from roughly 1 GB to 10 GB.
- Pro: Timestamped output and six export formats support subtitles, documents, and structured pipelines.
- Con: Results vary widely by language, so broad multilingual coverage does not ensure uniform accuracy.
- Con: Larger models demand substantial VRAM, and installation depends on ffmpeg with Rust potentially required for tiktoken.
- Con: Turbo cannot perform translation; non-English-to-English translation requires another multilingual model.
Alternatives
For evaluation-focused tooling rather than speech recognition, Maxim AI offers a freemium platform with API, self-hosted, and web options, including a free Developer plan with up to three seats, one workspace, 10,000 logs per month, and three-day retention.
Promptfoo is another freemium choice for LLM evaluation, available on Linux, macOS, Windows, web, API, and self-hosted deployments, with a free Community plan.
DeepEval is a free, Apache 2.0-licensed open-source LLM evaluation framework with local and CI/CD test running.
Giskard provides a freemium, locally deployable option for LLM vulnerability scanning and RAG evaluation.
Arena is the better fit for public, head-to-head model comparisons through a web leaderboard, with daily usage limits.
Braintrust suits teams wanting hosted evaluation workflows, with a free Starter plan covering 1 GB of processed data, 10,000 scores, 14-day retention, and unlimited users, projects, and datasets.
Galileo targets managed evaluation and analytics, with a Pro plan priced at $100 a month when billed yearly.
Parea AI offers a freemium evaluation platform; its free plan allows two team members, 3,000 logs per month with one-month retention, and 10 deployed prompts.
For broader discovery, compare speech recognition software, speech-to-text software, transcription software, AI video caption generators, and text-to-speech software.
Verdict
Choose Whisper if you need a free, MIT-licensed foundation that runs locally and spans multilingual transcription, English translation, language identification, and voice activity detection. Look elsewhere when your priority is consistently superior performance in a specific language, translation with the turbo model, or a managed service that removes installation and hardware responsibilities.
The verdict
Whisper provides a free, locally installable foundation for multilingual speech tasks, with several model sizes to choose from. Account for its installation requirements and language-dependent performance when deciding whether it fits your workflow.
Questions people ask
Is Whisper free?
Yes. The listed Whisper plan is free, and the project is open source.
Which platforms can run Whisper?
Whisper supports Linux, macOS, Windows, and self-hosted deployments.
What can Whisper export?
The repository lists TXT, VTT, SRT, TSV, JSON, and JSONL exports, with timestamp support.
Can Whisper translate speech?
It can translate speech from multiple languages into English. The turbo model is not trained for translation.
What license covers Whisper?
The code and model weights are released under the MIT License.
What the maker publishes
- API endpoints
- The model documentation lists transcription and translation capabilities and OpenAI API endpoint families including audio transcription. source
- API input handling
- The Audio API requires an uploaded file in a supported audio format and does not accept links to audio files. source
- API model
- OpenAI exposes whisper-1 as a general-purpose transcription model with audio input and text output. source
- API rate limits
- The documented Free tier allows 3 requests per minute and 200 requests per day, while Build allows 2,500 requests per minute. source
- API streaming limitation
- OpenAI states that streaming is not supported with the whisper-1 model. source
- API upload limit
- For legacy whisper-1 transcription uploads, the maximum request size is 25 MiB. source
- Automatic language identification
- Language identification is a built-in task alongside transcription and translation. source
- Company business
- Crunchbase describes OpenAI as developing artificial-intelligence models and deploying machine-learning software tools. source
- Company founding
- OpenAI was founded in December 2015. source
- Company headquarters and size
- LinkedIn lists OpenAI headquarters in San Francisco, California and company size of 1,001–5,000 employees. source
- Company legal name
- Crunchbase lists the legal name as OpenAI OpCo, LLC and also lists OpenAI LP and OpenAI Group PBC as alternate names. source
- Developer extensibility
- The repository points users to third-party web demos, integrations, platform ports and other extensions through GitHub Discussions. source
- Export formats
- txt,vtt,srt,tsv,json,jsonl source
- Hallucination criticism
- TechCrunch reported October 26, 2024 concerns from researchers that Whisper can insert invented words or passages, including in medical contexts. source
- Hardware
- The listed models require approximately 1 GB to 10 GB of VRAM, depending on model size. source
- Hardware requirements
- Approximate required VRAM ranges from about 1 GB for tiny and base to about 10 GB for large. source
- Independent accuracy test
- A September 2026 FreeTTS test measured 11.9% word error for large-v3 with voice-activity filtering and previous-text conditioning disabled, versus 14.8% on defaults. source
- Independent forensic limitation
- A Frontiers study found Whisper best among tested systems on poor-quality forensic-like audio but with about 50% word error, making it unsuitable for forensic use. source
- Installation
- The official installation command is pip install -U openai-whisper, with a Git installation option also documented. source
- Installation requirements
- The repository documents installation through pip and requires the ffmpeg command-line tool; Rust may be needed if a prebuilt tiktoken wheel is unavailable. source
- Integrations and examples
- The repository links a Google Colab example and invites users to share third-party extensions and integrations in GitHub Discussions. source
- Intended users
- OpenAI describes the open-sourced models and inference code as a foundation for building applications and further research on robust speech processing. source
- Language support
- Whisper supports multilingual speech recognition and translation into English, with the repository tokenizer listing available languages. source
- Licence
- Whisper code and model weights are released under the MIT License. source
Checked against the maker's pages on 21 Oct 2026.