Python tools for WhisperKit
- Convert PyTorch Whisper models to WhisperKit format
- Apply custom inference optimizations and model compression
- Evaluate Whisper using WhisperKit and other Whisper implementations on benchmarks
- Installation
- Model Generation
- Model Evaluation
- Python Inference
- Example SwiftUI App
- Quality-of-Inference
- FAQ
- Citation
- Step 1: Fork this repository
- Step 2: Create a Python virtual environment, e.g.:
conda create -n whisperkit python=3.11 -y && conda activate whisperkit- Step 3: Install as editable
cd WHISPERKIT_ROOT_DIR && pip install -e .Convert Hugging Face Whisper Models (PyTorch) to WhisperKit (Core ML) format:
whisperkit-generate-model --model-version <model-version> --output-dir <output-dir>For optional arguments related to model optimizations, please see the help menu with -h
We host several popular Whisper model versions here. These hosted models are automatically over-the-air deployable to apps integrating WhisperKit such as our example app WhisperAX on TestFlight. If you would like to publish custom Whisper versions that are not already published, you can do so as follows:
- Step 1: Find the user or organization name that you have write access to on Hugging Face Hub. If you are logged into
huggingface-clilocally, you may simply do:
huggingface-cli whoamiIf you don't have a write token yet, you can generate it here.
- Step 2: Point to the model repository that you would like to publish to, e.g.
my-org/my-whisper-repo-name, with theMODEL_REPO_IDenvironment variable and specify the name of the source PyTorch Whisper repository (e.g. distil-whisper/distil-small.en)
MODEL_REPO_ID=my-org/my-whisper-repo-name whisperkit-generate-model --model-version distil-whisper/distil-small.en --output-dir <output-dir>If the above command is successfuly executed, your model will have been published to hf.co/my-org/my-whisper-repo-name/distil-whisper_distil-small.en!
Evaluate (Argmax- or developer-published) models on speech recognition datasets:
whisperkit-evaluate-model --model-version <model-version> --output-dir <output-dir> --evaluation-dataset {librispeech-debug,librispeech,earnings22}By default, this command uses the latest main branch commits from WhisperKit and searches within Argmax-published model repositories. For optional arguments related to code and model versioning, please see the help menu with -h
We continually publish the evaluation results of Argmax-hosted models here as part of our continuous integration tests.
If you would like to evaluate WhisperKit models on your own dataset:
- Step 1: Publish a dataset on the Hub with the same simple structure as this toy dataset (audio files +
metadata.json) - Step 2: Run evaluation with environment variables as follows:
export CUSTOM_EVAL_DATASET="my-dataset-name-on-hub"
export DATASET_REPO_OWNER="my-user-or-org-name-on-hub"
export MODEL_REPO_ID="my-org/my-whisper-repo-name" # if evaluating self-published models
whisperkit-evaluate-model --model-version <model-version> --output-dir <output-dir> --evaluation-dataset my-dataset-name-on-hubUse the unified Python wrapper for several Whisper frameworks:
from whisperkit.pipelines import WhisperKit, WhisperCpp, WhisperMLX
pipe = WhisperKit(whisper_version="openai/whisper-large-v3", out_dir="/path/to/out/dir")
print(pipe("audio.{wav,flac,mp3}"))Note: Using WhisperCpp requires ffmpeg to be installed. Recommended installation is with brew install ffmpeg
This app serves two purposes:
- Base template for developers to freely customize and integrate parts into their own app
- Real-world testing/debugging utility for custom Whisper versions or WhisperKit features before/without building an app.
Note that the app is in beta and we are actively seeking feedback to improve it before widely distributing it.
Short-form Audio (<30s/clip) - 5 hours of English audiobook clips
| WER (↓) | QoI (↑) | File Size (MB) | Code Commit | |
|---|---|---|---|---|
| large-v2 (WhisperOpenAIAPI) | 2.35 | 100 | 3100 | N/A |
| large-v2 | 2.77 | 96.6 | 3100 | Link |
| large-v2_949MB | 2.4 | 94.6 | 949 | Link |
| large-v2_turbo | 2.76 | 96.6 | 3100 | Link |
| large-v2_turbo_955MB | 2.41 | 94.6 | 955 | Link |
| large-v3 | 2.04 | 95.2 | 3100 | Link |
| large-v3_947MB | 2.46 | 93.9 | 947 | Link |
| large-v3_turbo | 2.03 | 95.4 | 3100 | Link |
| large-v3_turbo_954MB | 2.47 | 93.9 | 954 | Link |
| distil-large-v3 | 2.47 | 89.7 | 1510 | Link |
| distil-large-v3_594MB | 2.96 | 85.4 | 594 | Link |
| distil-large-v3_turbo | 2.47 | 89.7 | 1510 | Link |
| distil-large-v3_turbo_600MB | 2.78 | 86.2 | 600 | Link |
| small.en | 3.12 | 85.8 | 483 | Link |
| small | 3.45 | 83 | 483 | Link |
| base.en | 3.98 | 75.3 | 145 | Link |
| base | 4.97 | 67.2 | 145 | Link |
| tiny.en | 5.61 | 63.9 | 66 | Link |
| tiny | 7.47 | 52.5 | 66 | Link |
Long-Form Audio (>1hr/clip) - 120 hours of earnings call recordings in English with various accents
| WER (↓) | QoI (↑) | File Size (MB) | Code Commit | |
|---|---|---|---|---|
| large-v2 (WhisperOpenAIAPI) | 16.27 | 100 | 3100 | N/A |
| large-v3 | 15.17 | 58.5 | 3100 | Link |
| base.en | 23.49 | 6.5 | 145 | Link |
| tiny.en | 28.64 | 5.7 | 66 | Link |
We believe that rigorously measuring the quality of inference is necessary for developers and
enterprises to make informed decisions when opting to use optimized or compressed variants of
any machine learning model in production. To contextualize WhisperKit, we take the following Whisper
implementations and benchmark them using a consistent evaluation harness:
Server-side:
WhisperOpenAIAPI: OpenAI's Whisper API
($0.36 per hour of audio as of 02/29/24, 25MB file size limit per request)
On-device:
WhisperKit: Argmax's implementation [Eval Harness] [Repo]whisper.cpp: A C++ implementation form ggerganov [Eval Harness] [Repo]WhisperMLX: A Python implementation from Apple MLX [Eval Harness] [Repo]
(All on-device implementations are available for free under MIT license as of 03/19/2024)
WhisperOpenAIAPI sets the reference and we assume that it is using the equivalent of openai/whisper-large-v2
in float16 precision along with additional undisclosed optimizations from OpenAI. In all measurements, we care primarily about per-example no-regressions (quantified as qoi below)
which is a stricter metric compared to dataset average Word Error RATE (WER). A 100% qoi preserves perfect backwards-compatibility on the test distribution and avoids "perceived regressions", the phenomenon
where per-example known behavior changes after a code/model update and causes divergence in downstream code or breaks the user experience itself (even if dataset averages might stay flat
across updates). Pseudocode for qoi:
qoi = []
for example in dataset:
no_regression = wer(optimized_model(example)) <= wer(reference_model(example))
qoi.append(no_regression)
qoi = (sum(qoi) / len(qoi)) * 100.Note that the ordering of models with respect to WER does not necessarily match the ordering with respect to QoI. This is because the reference model gets assigned
a QoI of 100% by definition. Any per-example regression by other implementations get penalized while per-example improvements are not rewarded. QoI (higher is better) matters
where the production behavior is established by the reference results and the goal is to not regress when switching to an optimized or compressed model. On the other hand,
WER (lower is better) matters when there is no established production behavior and one is picking the best quality versus model size trade off point.
We anticipate developers that use Whisper (or similar models) in production to have their own Quality Assurance test sets and whisperkittools offers the tooling necessary to run the same measurements on such custom test sets, please see the Model Evaluation on Custom Dataset for details.
WhisperKit is an SDK for building speech-to-text features in apps across a wide range of Apple devices. We are working towards abstracting away the model versioning from the developer so WhisperKit "just works" by deploying the highest-quality model version that a particular device can execute. In the interim, we leave the choice to the developer by providing quality and size trade-offs.
- librispeech: ~5 hours of short English audio clips, tests short-form transcription quality
- earnings22: ~120 hours of English audio clips from earnings calls with various accents, tests long-form transcription quality
Benchmark results on this page were automatically generated by whisperkittools using our cluster of Apple Silicon Macs as self-hosted runners on
Github Actions. We periodically recompute these benchmarks as part of our CI pipeline. Due to security concerns,
we are unable to open up the cluster to the public. However, any Apple Silicon Mac (even with 8GB RAM) can be used to
run identical evaluation jobs locally. For reference, our M2 Ultra devices complete a librispeech + openai/whisper-large-v3
evaluation in under 1 hour regardless of the Whisper implementation. Oldest Apple Silicon Macs should take less than 1 day to complete the same evaluation.
-
_turbo: Indicates the presence of additional optimizations (not compression) to unlock streaming transcription as described in our Blog Post. -
_*MB: Indicates the presence of model compression. Instead of cluttering the filename with details like_AudioEncoder-5.8bits_TextDecoder-6.1bits_QLoRA-rank=16, we choose to summarize the compression spec as the resulting total file size since this is what matters to developers in production.
Q1: xcrun: error: unable to find utility "coremlcompiler", not a developer tool or in PATH
A1: Ensure Xcode is installed on your Mac and run sudo xcode-select --switch /Applications/Xcode.app/Contents/Developer.
If you use WhisperKit for something cool or just find it useful, please drop us a note at [email protected]!
If you use WhisperKit for academic work, here is the BibTeX:
@misc{whisperkit-argmax,
title = {WhisperKit},
author = {Argmax, Inc.},
year = {2024},
URL = {https://github.com/argmaxinc/WhisperKit}
}
