From 25e5c364e0a21ddefee46adb674c591f1ba610ba Mon Sep 17 00:00:00 2001 From: Jong Wook Kim Date: Mon, 30 Sep 2024 10:59:51 -0700 Subject: [PATCH] large-v3-turbo model (#2361) --- README.md | 20 ++++++++++++-------- model-card.md | 4 +++- whisper/__init__.py | 4 ++++ whisper/transcribe.py | 2 +- 4 files changed, 20 insertions(+), 10 deletions(-) diff --git a/README.md b/README.md index afca9c971..910b7dbae 100644 --- a/README.md +++ b/README.md @@ -57,17 +57,21 @@ pip install setuptools-rust ## Available models and languages -There are five 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; actual speed may vary depending on many factors including the available hardware. +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 | ~32x | -| base | 74 M | `base.en` | `base` | ~1 GB | ~16x | -| small | 244 M | `small.en` | `small` | ~2 GB | ~6x | +| 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. @@ -77,9 +81,9 @@ Whisper's performance varies widely depending on the language. The figure below ## Command-line usage -The following command will transcribe speech in audio files, using the `medium` model: +The following command will transcribe speech in audio files, using the `turbo` model: - whisper audio.flac audio.mp3 audio.wav --model medium + whisper audio.flac audio.mp3 audio.wav --model turbo The default setting (which selects the `small` model) works well for transcribing English. To transcribe an audio file containing non-English speech, you can specify the language using the `--language` option: @@ -103,7 +107,7 @@ Transcription can also be performed within Python: ```python import whisper -model = whisper.load_model("base") +model = whisper.load_model("turbo") result = model.transcribe("audio.mp3") print(result["text"]) ``` @@ -115,7 +119,7 @@ Below is an example usage of `whisper.detect_language()` and `whisper.decode()` ```python import whisper -model = whisper.load_model("base") +model = whisper.load_model("turbo") # load audio and pad/trim it to fit 30 seconds audio = whisper.load_audio("audio.mp3") diff --git a/model-card.md b/model-card.md index 3c041a1c0..291bc4bb1 100644 --- a/model-card.md +++ b/model-card.md @@ -16,13 +16,15 @@ The Whisper models are trained for speech recognition and translation tasks, cap | small | 244 M | ✓ | ✓ | | medium | 769 M | ✓ | ✓ | | large | 1550 M | | ✓ | +| turbo | 798 M | | ✓ | In December 2022, we [released an improved large model named `large-v2`](https://github.com/openai/whisper/discussions/661), and `large-v3` in November 2023. +Additionally, we've added a `turbo` model in September 2024 which is optimized for inference speed. ### Release date -September 2022 (original series), December 2022 (`large-v2`), and November 2023 (`large-v3`) +September 2022 (original series), December 2022 (`large-v2`), November 2023 (`large-v3`), September 2024 (`large-v3-turbo`) ### Model type diff --git a/whisper/__init__.py b/whisper/__init__.py index d7fbba36f..e210718f3 100644 --- a/whisper/__init__.py +++ b/whisper/__init__.py @@ -27,6 +27,8 @@ "large-v2": "https://openaipublic.azureedge.net/main/whisper/models/81f7c96c852ee8fc832187b0132e569d6c3065a3252ed18e56effd0b6a73e524/large-v2.pt", "large-v3": "https://openaipublic.azureedge.net/main/whisper/models/e5b1a55b89c1367dacf97e3e19bfd829a01529dbfdeefa8caeb59b3f1b81dadb/large-v3.pt", "large": "https://openaipublic.azureedge.net/main/whisper/models/e5b1a55b89c1367dacf97e3e19bfd829a01529dbfdeefa8caeb59b3f1b81dadb/large-v3.pt", + "large-v3-turbo": "https://openaipublic.azureedge.net/main/whisper/models/aff26ae408abcba5fbf8813c21e62b0941638c5f6eebfb145be0c9839262a19a/large-v3-turbo.pt", + "turbo": "https://openaipublic.azureedge.net/main/whisper/models/aff26ae408abcba5fbf8813c21e62b0941638c5f6eebfb145be0c9839262a19a/large-v3-turbo.pt", } # base85-encoded (n_layers, n_heads) boolean arrays indicating the cross-attention heads that are @@ -44,6 +46,8 @@ "large-v2": b"ABzY8zd+h!0{>%R7=D0pU<_bnWW*tkYAhobTNnu$jnkEkXqp)j;w1Tzk)UH3X%SZd&fFZ2fC2yj", "large-v3": b"ABzY8gWO1E0{>%R7(9S+Kn!D~%ngiGaR?*L!iJG9p-nab0JQ=-{D1-g00", "large": b"ABzY8gWO1E0{>%R7(9S+Kn!D~%ngiGaR?*L!iJG9p-nab0JQ=-{D1-g00", + "large-v3-turbo": b"ABzY8j^C+e0{>%RARaKHP%t(lGR*)0g!tONPyhe`", + "turbo": b"ABzY8j^C+e0{>%RARaKHP%t(lGR*)0g!tONPyhe`", } diff --git a/whisper/transcribe.py b/whisper/transcribe.py index 1c075a201..8e1240bd6 100644 --- a/whisper/transcribe.py +++ b/whisper/transcribe.py @@ -511,7 +511,7 @@ def valid_model_name(name): # fmt: off parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter) parser.add_argument("audio", nargs="+", type=str, help="audio file(s) to transcribe") - parser.add_argument("--model", default="small", type=valid_model_name, help="name of the Whisper model to use") + parser.add_argument("--model", default="turbo", type=valid_model_name, help="name of the Whisper model to use") parser.add_argument("--model_dir", type=str, default=None, help="the path to save model files; uses ~/.cache/whisper by default") parser.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu", help="device to use for PyTorch inference") parser.add_argument("--output_dir", "-o", type=str, default=".", help="directory to save the outputs")