Whisper is a state-of-the-art model for automatic speech recognition (ASR) and speech translation, proposed in the paper Robust Speech Recognition via Large-Scale Weak Supervision, and Whisper large-v3-turbo is a finetuned version of a pruned Whisper large-v3. In other words, it's the exact same model, except that the number of decoding layers have reduced from 32 to 4. As a result, the model is way faster, at the expense of a minor quality degradation.

This model is a continued fine-tune using the dataset mozilla-foundation/common_voice_17_0, using the subset "es" and filtering out accents from Spain in order to improve performance over low-quality audios across latin-american accents.

Special thanks to CITECCA, in the Universidad Nacional de Rio Negro for providing the compute to train this model.

Evaluation

Evaluated against base model

--- Eval Results ---
Comparing 'openai/whisper-large-v3-turbo' with 'marianbasti/whisper-large-v3-turbo-latam/' on 3952 samples of "mozilla-foundation/common_voice_17_0" (Spanish subset, evaluation split).

--- Word error rate ---
  Default model (openai/whisper-large-v3-turbo): 15.44%
  Finetuned model (marianbasti/whisper-large-v3-turbo-lataml):  7.80%
  Improvement: 7.64% (better)

--- Average inference time ---
  Default model: 0.56 seconds per sample
  Custom model:  0.34 seconds per sample
  Speed improvement: 1.62x faster

Usage

This model is supported in Hugging Face 🤗 Transformers. To run the model, first install the Transformers library. For this example, we'll also install 🤗 Datasets to load toy audio dataset from the Hugging Face Hub, and 🤗 Accelerate to reduce the model loading time:

bash

pip install --upgrade pip
pip install --upgrade transformers datasets[audio] accelerate

The model can be used with the pipeline class to transcribe audios of arbitrary length:

python

import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
from datasets import load_dataset


device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model_id = "marianbasti/whisper-large-v3-turbo-latam"

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True
)
model.to(device)

processor = AutoProcessor.from_pretrained(model_id)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    torch_dtype=torch_dtype,
    device=device,
)

dataset = load_dataset("distil-whisper/librispeech_long", "clean", split="validation")
sample = dataset[0]["audio"]

result = pipe(sample)
print(result["text"])

To transcribe a local audio file, simply pass the path to your audio file when you call the pipeline:

result = pipe("audio.mp3")

Multiple audio files can be transcribed in parallel by specifying them as a list and setting the batch_size parameter:

result = pipe(["audio_1.mp3", "audio_2.mp3"], batch_size=2)

Transformers is compatible with all Whisper decoding strategies, such as temperature fallback and condition on previous tokens. The following example demonstrates how to enable these heuristics:

python

generate_kwargs = {
    "max_new_tokens": 448,
    "num_beams": 1,
    "condition_on_prev_tokens": False,
    "compression_ratio_threshold": 1.35,  # zlib compression ratio threshold (in token space)
    "temperature": (0.0, 0.2, 0.4, 0.6, 0.8, 1.0),
    "logprob_threshold": -1.0,
    "no_speech_threshold": 0.6,
    "return_timestamps": True,
}

result = pipe(sample, generate_kwargs=generate_kwargs)

Whisper predicts the language of the source audio automatically. If the source audio language is known a-priori, it can be passed as an argument to the pipeline:

result = pipe(sample, generate_kwargs={"language": "english"})

By default, Whisper performs the task of speech transcription, where the source audio language is the same as the target text language. To perform speech translation, where the target text is in English, set the task to "translate":

result = pipe(sample, generate_kwargs={"task": "translate"})

Finally, the model can be made to predict timestamps. For sentence-level timestamps, pass the return_timestamps argument:

python

result = pipe(sample, return_timestamps=True)
print(result["chunks"])

And for word-level timestamps:

python

result = pipe(sample, return_timestamps="word")
print(result["chunks"])

The above arguments can be used in isolation or in combination. For example, to perform the task of speech transcription where the source audio is in French, and we want to return sentence-level timestamps, the following can be used:

python

result = pipe(sample, return_timestamps=True, generate_kwargs={"language": "french", "task": "translate"})
print(result["chunks"])

For more control over the generation parameters, use the model + processor API directly:

python

import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
from datasets import Audio, load_dataset


device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model_id = "marianbasti/whisper-large-v3-turbo-latam"

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True
)
model.to(device)

processor = AutoProcessor.from_pretrained(model_id)

dataset = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
dataset = dataset.cast_column("audio", Audio(processor.feature_extractor.sampling_rate))
sample = dataset[0]["audio"]

inputs = processor(
    sample["array"],
    sampling_rate=sample["sampling_rate"],
    return_tensors="pt",
    truncation=False,
    padding="longest",
    return_attention_mask=True,
)
inputs = inputs.to(device, dtype=torch_dtype)

gen_kwargs = {
    "max_new_tokens": 448,
    "num_beams": 1,
    "condition_on_prev_tokens": False,
    "compression_ratio_threshold": 1.35,  # zlib compression ratio threshold (in token space)
    "temperature": (0.0, 0.2, 0.4, 0.6, 0.8, 1.0),
    "logprob_threshold": -1.0,
    "no_speech_threshold": 0.6,
    "return_timestamps": True,
}

pred_ids = model.generate(**inputs, **gen_kwargs)
pred_text = processor.batch_decode(pred_ids, skip_special_tokens=True, decode_with_timestamps=False)

print(pred_text)

Additional Speed & Memory Improvements

You can apply additional speed and memory improvements to Whisper to further reduce the inference speed and VRAM requirements.

Chunked Long-Form

Whisper has a receptive field of 30-seconds. To transcribe audios longer than this, one of two long-form algorithms are required:

Sequential: uses a "sliding window" for buffered inference, transcribing 30-second slices one after the other
Chunked: splits long audio files into shorter ones (with a small overlap between segments), transcribes each segment independently, and stitches the resulting transcriptions at the boundaries

The sequential long-form algorithm should be used in either of the following scenarios:

Transcription accuracy is the most important factor, and speed is less of a consideration
You are transcribing batches of long audio files, in which case the latency of sequential is comparable to chunked, while being up to 0.5% WER more accurate

Conversely, the chunked algorithm should be used when:

Transcription speed is the most important factor
You are transcribing a single long audio file

By default, Transformers uses the sequential algorithm. To enable the chunked algorithm, pass the chunk_length_s parameter to the pipeline. For large-v3, a chunk length of 30-seconds is optimal. To activate batching over long audio files, pass the argument batch_size:

python

import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
from datasets import load_dataset


device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model_id = "marianbasti/whisper-large-v3-turbo-latam"

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True
)
model.to(device)

processor = AutoProcessor.from_pretrained(model_id)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    chunk_length_s=30,
    batch_size=16,  # batch size for inference - set based on your device
    torch_dtype=torch_dtype,
    device=device,
)

dataset = load_dataset("distil-whisper/librispeech_long", "clean", split="validation")
sample = dataset[0]["audio"]

result = pipe(sample)
print(result["text"])

Torch compile

The Whisper forward pass is compatible with torch.compile for 4.5x speed-ups.

Note: torch.compile is currently not compatible with the Chunked long-form algorithm or Flash Attention 2 ⚠️

python

import torch
from torch.nn.attention import SDPBackend, sdpa_kernel
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
from datasets import load_dataset
from tqdm import tqdm

torch.set_float32_matmul_precision("high")

device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model_id = "marianbasti/whisper-large-v3-turbo-latam"

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True
).to(device)

# Enable static cache and compile the forward pass
model.generation_config.cache_implementation = "static"
model.generation_config.max_new_tokens = 256
model.forward = torch.compile(model.forward, mode="reduce-overhead", fullgraph=True)

processor = AutoProcessor.from_pretrained(model_id)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    torch_dtype=torch_dtype,
    device=device,
)

dataset = load_dataset("distil-whisper/librispeech_long", "clean", split="validation")
sample = dataset[0]["audio"]

# 2 warmup steps
for _ in tqdm(range(2), desc="Warm-up step"):
    with sdpa_kernel(SDPBackend.MATH):
        result = pipe(sample.copy(), generate_kwargs={"min_new_tokens": 256, "max_new_tokens": 256})

# fast run
with sdpa_kernel(SDPBackend.MATH):
    result = pipe(sample.copy())

print(result["text"])

Flash Attention 2

We recommend using Flash-Attention 2 if your GPU supports it and you are not using torch.compile. To do so, first install Flash Attention:

pip install flash-attn --no-build-isolation

Then pass attn_implementation="flash_attention_2" to from_pretrained:

model = AutoModelForSpeechSeq2Seq.from_pretrained(model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, attn_implementation="flash_attention_2")

Torch Scale-Product-Attention (SDPA)

If your GPU does not support Flash Attention, we recommend making use of PyTorch scaled dot-product attention (SDPA). This attention implementation is activated by default for PyTorch versions 2.1.1 or greater. To check whether you have a compatible PyTorch version, run the following Python code snippet:

python

from transformers.utils import is_torch_sdpa_available

print(is_torch_sdpa_available())

If the above returns True, you have a valid version of PyTorch installed and SDPA is activated by default. If it returns False, you need to upgrade your PyTorch version according to the official instructions

Once a valid PyTorch version is installed, SDPA is activated by default. It can also be set explicitly by specifying attn_implementation="sdpa" as follows:

model = AutoModelForSpeechSeq2Seq.from_pretrained(model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, attn_implementation="sdpa")

For more information about how to use the SDPA refer to the Transformers SDPA documentation.

Evaluated Use

The primary intended users of these models are AI researchers studying robustness, generalization, capabilities, biases, and constraints of the current model. However, Whisper is also potentially quite useful as an ASR solution for developers, especially for English speech recognition. We recognize that once models are released, it is impossible to restrict access to only “intended” uses or to draw reasonable guidelines around what is or is not research.

The models are primarily trained and evaluated on ASR and speech translation to English tasks. They show strong ASR results in ~10 languages. They may exhibit additional capabilities, particularly if fine-tuned on certain tasks like voice activity detection, speaker classification, or speaker diarization but have not been robustly evaluated in these areas. We strongly recommend that users perform robust evaluations of the models in a particular context and domain before deploying them.

In particular, we caution against using Whisper models to transcribe recordings of individuals taken without their consent or purporting to use these models for any kind of subjective classification. We recommend against use in high-risk domains like decision-making contexts, where flaws in accuracy can lead to pronounced flaws in outcomes. The models are intended to transcribe and translate speech, use of the model for classification is not only not evaluated but also not appropriate, particularly to infer human attributes.

Training Data

No information provided.

BibTeX entry and citation info

bibtex

@misc{radford2022whisper,
  doi = {10.48550/ARXIV.2212.04356},
  url = {https://arxiv.org/abs/2212.04356},
  author = {Radford, Alec and Kim, Jong Wook and Xu, Tao and Brockman, Greg and McLeavey, Christine and Sutskever, Ilya},
  title = {Robust Speech Recognition via Large-Scale Weak Supervision},
  publisher = {arXiv},
  year = {2022},
  copyright = {arXiv.org perpetual, non-exclusive license}
}

surus-ai/whisper-large-v3-turbo-latam

Whisper

Evaluation

Usage

Additional Speed & Memory Improvements

Chunked Long-Form

Torch compile

Flash Attention 2

Torch Scale-Product-Attention (SDPA)

Evaluated Use

Training Data

BibTeX entry and citation info

surus-ai/
whisper-large-v3-turbo-latam