## File: README.md # Whisper [[Blog]](https://openai.com/blog/whisper) [[Paper]](https://arxiv.org/abs/2212.04356) [[Model card]](https://github.com/openai/whisper/blob/main/model-card.md) [[Colab example]](https://colab.research.google.com/github/openai/whisper/blob/master/notebooks/LibriSpeech.ipynb) Whisper is a general-purpose speech recognition model. It is trained on a large dataset of diverse audio and is also a multitasking model that can perform multilingual speech recognition, speech translation, and language identification. ## Approach A Transformer sequence-to-sequence model is trained on various speech processing tasks, including multilingual speech recognition, speech translation, spoken language identification, and voice activity detection. These tasks are jointly represented as a sequence of tokens to be predicted by the decoder, allowing a single model to replace many stages of a traditional speech-processing pipeline. The multitask training format uses a set of special tokens that serve as task specifiers or classification targets. ## Setup We used Python 3.9.9 and [PyTorch](https://pytorch.org/) 1.10.1 to train and test our models, but the codebase is expected to be compatible with Python 3.8-3.11 and recent PyTorch versions. The codebase also depends on a few Python packages, most notably [OpenAI's tiktoken](https://github.com/openai/tiktoken) for their fast tokenizer implementation. You can download and install (or update to) the latest release of Whisper with the following command: pip install -U openai-whisper Alternatively, the following command will pull and install the latest commit from this repository, along with its Python dependencies: pip install git+https://github.com/openai/whisper.git To update the package to the latest version of this repository, please run: pip install --upgrade --no-deps --force-reinstall git+https://github.com/openai/whisper.git It also requires the command-line tool [`ffmpeg`](https://ffmpeg.org/) to be installed on your system, which is available from most package managers: ```bash # on Ubuntu or Debian sudo apt update && sudo apt install ffmpeg # on Arch Linux sudo pacman -S ffmpeg # on MacOS using Homebrew (https://brew.sh/) brew install ffmpeg # on Windows using Chocolatey (https://chocolatey.org/) choco install ffmpeg # on Windows using Scoop (https://scoop.sh/) scoop install ffmpeg ``` You may need [`rust`](http://rust-lang.org) installed as well, in case [tiktoken](https://github.com/openai/tiktoken) does not provide a pre-built wheel for your platform. If you see installation errors during the `pip install` command above, please follow the [Getting started page](https://www.rust-lang.org/learn/get-started) to install Rust development environment. Additionally, you may need to configure the `PATH` environment variable, e.g. `export PATH="$HOME/.cargo/bin:$PATH"`. If the installation fails with `No module named 'setuptools_rust'`, you need to install `setuptools_rust`, e.g. by running: ```bash pip install setuptools-rust ``` ## Available models and languages There are six model sizes, four with English-only versions, offering speed and accuracy tradeoffs. Below are the names of the available models and their approximate memory requirements and inference speed relative to the large model. The relative speeds below are measured by transcribing English speech on a A100, and the real-world speed may vary significantly depending on many factors including the language, the speaking speed, and the available hardware. | Size | Parameters | English-only model | Multilingual model | Required VRAM | Relative speed | |:------:|:----------:|:------------------:|:------------------:|:-------------:|:--------------:| | tiny | 39 M | `tiny.en` | `tiny` | ~1 GB | ~10x | | base | 74 M | `base.en` | `base` | ~1 GB | ~7x | | small | 244 M | `small.en` | `small` | ~2 GB | ~4x | | medium | 769 M | `medium.en` | `medium` | ~5 GB | ~2x | | large | 1550 M | N/A | `large` | ~10 GB | 1x | | turbo | 809 M | N/A | `turbo` | ~6 GB | ~8x | The `.en` models for English-only applications tend to perform better, especially for the `tiny.en` and `base.en` models. We observed that the difference becomes less significant for the `small.en` and `medium.en` models. Additionally, the `turbo` model is an optimized version of `large-v3` that offers faster transcription speed with a minimal degradation in accuracy. Whisper's performance varies widely depending on the language. The figure below shows a performance breakdown of `large-v3` and `large-v2` models by language, using WERs (word error rates) or CER (character error rates, shown in *Italic*) evaluated on the Common Voice 15 and Fleurs datasets. Additional WER/CER metrics corresponding to the other models and datasets can be found in Appendix D.1, D.2, and D.4 of [the paper](https://arxiv.org/abs/2212.04356), as well as the BLEU (Bilingual Evaluation Understudy) scores for translation in Appendix D.3. ## Command-line usage The following command will transcribe speech in audio files, using the `turbo` model: ```bash whisper audio.flac audio.mp3 audio.wav --model turbo ``` The default setting (which selects the `turbo` model) works well for transcribing English. However, **the `turbo` model is not trained for translation tasks**. If you need to **translate non-English speech into English**, use one of the **multilingual models** (`tiny`, `base`, `small`, `medium`, `large`) instead of `turbo`. For example, to transcribe an audio file containing non-English speech, you can specify the language: ```bash whisper japanese.wav --language Japanese ``` To **translate** speech into English, use: ```bash whisper japanese.wav --model medium --language Japanese --task translate ``` > **Note:** The `turbo` model will return the original language even if `--task translate` is specified. Use `medium` or `large` for the best translation results. Run the following to view all available options: ```bash whisper --help ``` See [tokenizer.py](https://github.com/openai/whisper/blob/main/whisper/tokenizer.py) for the list of all available languages. ## Python usage Transcription can also be performed within Python: ```python import whisper model = whisper.load_model("turbo") result = model.transcribe("audio.mp3") print(result["text"]) ``` Internally, the `transcribe()` method reads the entire file and processes the audio with a sliding 30-second window, performing autoregressive sequence-to-sequence predictions on each window. Below is an example usage of `whisper.detect_language()` and `whisper.decode()` which provide lower-level access to the model. ```python import whisper model = whisper.load_model("turbo") # load audio and pad/trim it to fit 30 seconds audio = whisper.load_audio("audio.mp3") audio = whisper.pad_or_trim(audio) # make log-Mel spectrogram and move to the same device as the model mel = whisper.log_mel_spectrogram(audio, n_mels=model.dims.n_mels).to(model.device) # detect the spoken language _, probs = model.detect_language(mel) print(f"Detected language: {max(probs, key=probs.get)}") # decode the audio options = whisper.DecodingOptions() result = whisper.decode(model, mel, options) # print the recognized text print(result.text) ``` ## More examples Please use the [🙌 Show and tell](https://github.com/openai/whisper/discussions/categories/show-and-tell) category in Discussions for sharing more example usages of Whisper and third-party extensions such as web demos, integrations with other tools, ports for different platforms, etc. ## License Whisper's code and model weights are released under the MIT License. See [LICENSE](https://github.com/openai/whisper/blob/main/LICENSE) for further details. --- ## File: data/README.md This directory supplements the paper with more details on how we prepared the data for evaluation, to help replicate our experiments. ## Short-form English-only datasets ### LibriSpeech We used the test-clean and test-other splits from the [LibriSpeech ASR corpus](https://www.openslr.org/12). ### TED-LIUM 3 We used the test split of [TED-LIUM Release 3](https://www.openslr.org/51/), using the segmented manual transcripts included in the release. ### Common Voice 5.1 We downloaded the English subset of Common Voice Corpus 5.1 from [the official website](https://commonvoice.mozilla.org/en/datasets) ### Artie We used the [Artie bias corpus](https://github.com/artie-inc/artie-bias-corpus). This is a subset of the Common Voice dataset. ### CallHome & Switchboard We used the two corpora from [LDC2002S09](https://catalog.ldc.upenn.edu/LDC2002S09) and [LDC2002T43](https://catalog.ldc.upenn.edu/LDC2002T43) and followed the [eval2000_data_prep.sh](https://github.com/kaldi-asr/kaldi/blob/master/egs/fisher_swbd/s5/local/eval2000_data_prep.sh) script for preprocessing. The `wav.scp` files can be converted to WAV files with the following bash commands: ```bash mkdir -p wav while read name cmd; do echo $name echo ${cmd/\|/} wav/$name.wav | bash done < wav.scp ``` ### WSJ We used [LDC93S6B](https://catalog.ldc.upenn.edu/LDC93S6B) and [LDC94S13B](https://catalog.ldc.upenn.edu/LDC94S13B) and followed the [s5 recipe](https://github.com/kaldi-asr/kaldi/tree/master/egs/wsj/s5) to preprocess the dataset. ### CORAAL We used the 231 interviews from [CORAAL (v. 2021.07)](https://oraal.uoregon.edu/coraal) and used the segmentations from [the FairSpeech project](https://github.com/stanford-policylab/asr-disparities/blob/master/input/CORAAL_transcripts.csv). ### CHiME-6 We downloaded the [CHiME-5 dataset](https://spandh.dcs.shef.ac.uk//chime_challenge/CHiME5/download.html) and followed the stage 0 of the [s5_track1 recipe](https://github.com/kaldi-asr/kaldi/tree/master/egs/chime6/s5_track1) to create the CHiME-6 dataset which fixes synchronization. We then used the binaural recordings (`*_P??.wav`) and the corresponding transcripts. ### AMI-IHM, AMI-SDM1 We preprocessed the [AMI Corpus](https://groups.inf.ed.ac.uk/ami/corpus/overview.shtml) by following the stage 0 and 2 of the [s5b recipe](https://github.com/kaldi-asr/kaldi/tree/master/egs/ami/s5b). ## Long-form English-only datasets ### TED-LIUM 3 To create a long-form transcription dataset from the [TED-LIUM3](https://www.openslr.org/51/) dataset, we sliced the audio between the beginning of the first labeled segment and the end of the last labeled segment of each talk, and we used the concatenated text as the label. Below are the timestamps used for slicing each of the 11 TED talks in the test split. | Filename | Begin time (s) | End time (s) | |---------------------|----------------|--------------| | DanBarber_2010 | 16.09 | 1116.24 | | JaneMcGonigal_2010 | 15.476 | 1187.61 | | BillGates_2010 | 15.861 | 1656.94 | | TomWujec_2010U | 16.26 | 402.17 | | GaryFlake_2010 | 16.06 | 367.14 | | EricMead_2009P | 18.434 | 536.44 | | MichaelSpecter_2010 | 16.11 | 979.312 | | DanielKahneman_2010 | 15.8 | 1199.44 | | AimeeMullins_2009P | 17.82 | 1296.59 | | JamesCameron_2010 | 16.75 | 1010.65 | | RobertGupta_2010U | 16.8 | 387.03 | ### Meanwhile This dataset consists of 64 segments from The Late Show with Stephen Colbert. The YouTube video ID, start and end timestamps, and the labels can be found in [meanwhile.json](meanwhile.json). The labels are collected from the closed-caption data for each video and corrected with manual inspection. ### Rev16 We use a subset of 16 files from the 30 podcast episodes in [Rev.AI's Podcast Transcription Benchmark](https://www.rev.ai/blog/podcast-transcription-benchmark-part-1/), after finding that there are multiple cases where a significant portion of the audio and the labels did not match, mostly on the parts introducing the sponsors. We selected 16 episodes that do not have this error, whose "file number" are: 3 4 9 10 11 14 17 18 20 21 23 24 26 27 29 32 ### Kincaid46 This dataset consists of 46 audio files and the corresponding transcripts compiled in the blog article [Which automatic transcription service is the most accurate - 2018](https://medium.com/descript/which-automatic-transcription-service-is-the-most-accurate-2018-2e859b23ed19) by Jason Kincaid. We used the 46 audio files and reference transcripts from the Airtable widget in the article. For the human transcription benchmark in the paper, we use a subset of 25 examples from this data, whose "Ref ID" are: 2 4 5 8 9 10 12 13 14 16 19 21 23 25 26 28 29 30 33 35 36 37 42 43 45 ### Earnings-21, Earnings-22 For these datasets, we used the files available in [the speech-datasets repository](https://github.com/revdotcom/speech-datasets), as of their `202206` version. ### CORAAL We used the 231 interviews from [CORAAL (v. 2021.07)](https://oraal.uoregon.edu/coraal) and used the full-length interview files and transcripts. ## Multilingual datasets ### Multilingual LibriSpeech We used the test splits from each language in [the Multilingual LibriSpeech (MLS) corpus](https://www.openslr.org/94/). ### Fleurs We collected audio files and transcripts using the implementation available as [HuggingFace datasets](https://huggingface.co/datasets/google/fleurs/blob/main/fleurs.py). To use as a translation dataset, we matched the numerical utterance IDs to find the corresponding transcript in English. ### VoxPopuli We used the `get_asr_data.py` script from [the official repository](https://github.com/facebookresearch/voxpopuli) to collect the ASR data in 14 languages. ### Common Voice 9 We downloaded the Common Voice Corpus 9 from [the official website](https://commonvoice.mozilla.org/en/datasets) ### CoVOST 2 We collected the `X into English` data collected using [the official repository](https://github.com/facebookresearch/covost).