Six model sizes cover speech, not speaker identity
Whisper's README presents 6 size classes, from tiny through turbo, for local speech recognition. The multilingual models can transcribe spoken language, identify a language, or translate speech into English. Four sizes also have English-only variants. The Python interface can load a model and return text, segments, timestamps, and language information, while the command-line tool writes common transcript and subtitle formats. Code and model weights use the MIT license, which keeps local and commercial integration straightforward.
Its scope is narrower than a finished meeting service. The model card says speaker diarization and speaker classification have not been thoroughly evaluated, and real-time transcription does not work out of the box. Whisper reads the entire audio file and processes it with a sliding 30-second window. Applications still need upload limits, job control, streaming or chunk assembly, speaker labeling, storage, and transcript review. For recorded interviews or searchable archives, that is manageable. For live multi-speaker captions, it is substantial missing product work.
A 30-second window can still invent or repeat text
OpenAI's model card says Whisper can generate words that were not spoken because it learned from weakly supervised, noisy data. It can also repeat text. Accuracy varies by language, accent, dialect, and demographic group, with lower-resource languages often faring worse. These are model limits rather than ordinary parser bugs. A valid JSON response or subtitle file proves only that decoding completed. It does not establish that names, numbers, negations, or specialized terms match the recording.
The practical response is domain testing and review. Build an evaluation set from your microphones, rooms, speakers, and vocabulary. Preserve timestamps and audio so a reviewer can check consequential passages. The model card advises against high-risk decision contexts and transcribing people without consent. It also says Whisper is not appropriate for inferring human attributes. That rules out using raw transcripts as unquestioned evidence in medicine, law, employment, or identity judgments, even when a larger checkpoint sounds fluent.
What happened when we ran it
Our sandbox installed commit 8609812 in 79 seconds, adding 81 Python packages and consuming 5,692 MB on disk. The repository checkout itself contained 45 files, about 4,315 lines of source, and used 10.4 MB. The build succeeded in 12 seconds. Pip-audit reported 0 known vulnerabilities in the installed environment. That is a clean audit result for the dependencies observed by that run, though it is not a security review of model files, FFmpeg inputs, or an application built around Whisper.
Tests did not finish successfully. The command ran for 563 seconds and exited with code 137. Its final output showed Python's threading stack inside tqdm's monitor wait, followed by pytest's timeout banner and an F. The supplied tail has no completed summary and does not name the underlying reason for exit 137, so we cannot report a test count or call it an out-of-memory failure. The precise result is that the suite failed after more than 9 minutes in the stated 3-CPU, 8 GB container.
The 5,692 MB install comes before model storage
Whisper has a one-line pip install, but the runtime is not a one-package utility. Our environment reached 5,692 MB after installing 81 packages. Audio decoding also requires the FFmpeg command-line program. On platforms without a prebuilt tiktoken wheel, the README says Rust and possibly setuptools-rust may be needed. The first useful run must obtain a model checkpoint, so image size, cache placement, network access, and cold-start policy all belong in a deployment plan.
Model choice changes hardware demand. The README estimates about 1 GB of VRAM for tiny and base, around 2 GB for small, 5 GB for medium, 6 GB for turbo, and 10 GB for large. Those are upstream approximations, and it warns that real speed varies by language, speaking rate, and hardware. CPU use is supported, but no single table predicts your turnaround. Time representative files on the actual host before promising delivery times or worker capacity.
Turbo cannot translate speech into English
The default CLI example selects turbo, an 809M-parameter model in the current README table. It is intended for fast transcription, but the README explicitly says it was not trained for translation. Supplying --task translate with turbo returns the original language rather than English. Users who need speech translation must choose a multilingual tiny, base, small, medium, or large model. This is easy to configure once and easy to miss when a generic wrapper always accepts the same task option.
Automatic language selection has another open edge. Pull request 2848 reports that clip_timestamps can start later in a file while language detection still samples up to the first 30 seconds from the beginning. Current source performs detection before parsing the requested clips, which supports the report. A multilingual recording with an excluded introduction can therefore choose the wrong decoding language. Setting the language explicitly avoids that path when the caller already knows it.
August code activity outweighs the 2025 release tag
GitHub recorded the last push on August 31, 2026, and fresh pull requests arrived in September. The repository had 108,596 stars and 145 combined issues and pull requests when fetched. Its latest tagged GitHub release was v20250625, published June 26, 2025. That older tag alone is poor evidence of abandonment because the default branch and review queue are current. It does mean package buyers should decide whether they want the tagged release or a newer commit and pin that choice.
Whisper remains a sensible baseline because its model limits are documented and its reference interface is small. Our successful 12-second build and 0-vulnerability audit are positives; the 5,692 MB installation and failed 563-second test run are real costs. Use it for local batch transcription when you can test the domain, inspect important passages, and operate the surrounding job system. Faster runtimes, native ports, and alignment pipelines are better buys when those surrounding needs define the product.