AI modelsOpenAI (weights); whisper.cpp/faster-whisper (community runtimes)
Whisper (open source)
Decision matrix
Whisper (open source) model decision matrix
Overview
OpenAI's open-source Whisper is a free, MIT-licensed speech recognition model in six sizes, from Tiny (39M parameters, ~1GB VRAM) to Large (1550M parameters, ~10GB VRAM), plus a Turbo variant balancing speed and accuracy for transcription. Pick the smallest size that's accurate enough for your task; you install and run command-line software, not a polished transcription app.
Other good fits
Multilingual transcription, translation and language identification from one model family.
+ Pros
Read pros
Whisper is OpenAI's open-source, general-purpose speech recognition model: multilingual transcription, translation, and language identification from one Transformer-based model family, released under the MIT license (both code and weights), so commercial and private use is broadly permitted. It comes in six sizes with real speed and accuracy trade-offs, letting you pick based on your hardware and task.
Pick it when…
You want free, self-hosted transcription and can pick a model size matching your actual hardware.
You need multilingual support, translation or language identification from the same model.
You're comfortable running command-line software rather than needing a polished GUI app.
− Cons
Read cons
Smaller models trade accuracy for speed and lower VRAM; don't assume Tiny or Base will match Large's accuracy on difficult audio (accents, background noise, technical vocabulary), test your actual use case at your chosen size. Even the largest model can hallucinate, generating text that wasn't actually spoken, so OpenAI's own model card cautions against using it for subjective classification or high-risk decision-making contexts. The model card also documents disparate performance across accents, dialects and lower-resource languages, and explicitly warns against transcribing anyone's recordings without their consent. This is genuinely command-line software: if you need a polished transcription app with a GUI, you'll need to either build one around Whisper or find a packaged tool that wraps it, since the raw project doesn't provide that.
Pick something else when…
You want a polished, managed transcription API instead of self-hosting.Deepgram is a comparable hosted speech-to-text option.
You need it as part of a realtime voice pipeline, not batch transcription.Pipecat or a realtime speech API fits that different use case better.
You want the fastest possible inference and are willing to pay for a specialized inference provider.Groq offers fast inference for compatible models.
Avoid if
You need translation quality specifically; the Turbo variant is optimized for transcription, not translation.
Context
English-only variants (with a .en suffix) give better accuracy for English-specific applications, particularly at smaller sizes; use those if you know your audio is English-only rather than the multilingual variant.
Requirements
Model
Parameters
Relative speed
VRAM
Tiny
39M
~10x faster
~1GB
Base
74M
~7x faster
~1GB
Small
244M
~4x faster
~2GB
Medium
769M
~2x faster
~5GB
Large
1550M
Standard (baseline)
~10GB
Turbo
809M
~8x faster
~6GB
Turbo is optimized for transcription specifically, not translation; pick Large or Medium if translation quality matters more than speed.
Licence & cost
The code and weights are free under MIT. You pay for the machine: VRAM requirements range from about 1GB (Tiny/Base) to about 10GB (Large), so pick a size your hardware can actually run at acceptable speed. Community runtimes like whisper.cpp and faster-whisper often give better performance than the reference implementation on the same hardware, so check those if speed matters. This is command-line software, not a polished app: plan for installation and configuration effort, not a one-click experience.
First task
Documentation-based suggestion — not run here.
Read first task
Transcribe a short recording and check accuracy against the source.
Install Whisper (or a community runtime like whisper.cpp or faster-whisper for better performance on your hardware).
Pick a model size matching your GPU's VRAM, starting with Small or Medium as a reasonable middle ground.
Transcribe a short recording (a few minutes) where you know exactly what was said.
Compare the transcript against the source, counting errors, especially around names, numbers and technical terms.
Done when you've compared a transcript against known-correct source audio and have a concrete sense of the error rate at your chosen model size, not just a vague impression of "pretty good."
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Rests on the OpenAI Whisper GitHub repository, checked 23 September 2026. No installation, transcription or benchmark was performed for this review.
Whisper repository · Sep 23, 2026 — Six model sizes with parameter counts, relative speed and VRAM figures, MIT license, English-only variants, Turbo trade-offs.
Transcribe lectures, interviews, and recordings with an openly released speech model.
Run whisper.cpp on CPUs or Apple Silicon, or use faster-whisper with a GPU.
Keep transcription local when your workflow requires offline processing.
