pretty_name: WhisperKit ASR Evaluation Results
viewer: false
library_name: whisperkit
tags:
- whisper
- whisperkit
- coreml
- asr
- quantized
Whisper Transcription Quality
Dataset: librispeech
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-whisper_distil-large-v3 | 2.47 | 89.7 | 1510 | Link |
| distil-whisper_distil-large-v3_594MB | 2.96 | 85.4 | 594 | Link |
| distil-whisper_distil-large-v3_turbo | 2.47 | 89.7 | 1510 | Link |
| distil-whisper_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 |
Dataset: earnings22
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 |
Explanation
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.
Why are there so many Whisper versions?
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.
Datasets
- 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
Reproducing Results
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.
Glossary
_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.