language:

• en
• fr
• es
• de license: mit library_name: transformers tags:
• audio
• automatic-speech-recognition
• transformers.js widget:
• example_title: LibriSpeech sample 1 src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
• example_title: LibriSpeech sample 2 src: https://cdn-media.huggingface.co/speech_samples/sample2.flac pipeline_tag: automatic-speech-recognition

Whisper-Large-V3-Distil-Multi4-v0.2

A multilingual distilled Whisper model with 2 decoder layers, supporting 4 European languages: English, French, Spanish, and German.

The model was trained during my work on Distil-Large-v3.5.

A notable feature is its native support for code-switching. The model has the ability to switch languages within a single segment transcription by automatically producing a new language token when it detects a language change (as demonstrated in the following example).

The <|yue|> language token has been repurposed during training to act as an automatic language detection token that enables code-switching during inference. To use this feature, simply set the language parameter to cantonese (used by default).

The model's performance is below both the monolingual distilled version and Whisper-Large-v3-Turbo. Future work should investigate better training procedures and possibly incorporate more data to effectively compress multilingual capabilities into a single model.

Table of Contents

• Usage
• Evaluation

Usage

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

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

# Load model
model_name_or_path = "bofenghuang/whisper-large-v3-distil-multi4-v0.2"
processor = AutoProcessor.from_pretrained(model_name_or_path)
model = AutoModelForSpeechSeq2Seq.from_pretrained(model_name_or_path, torch_dtype=torch_dtype)
model.to(device)

# Example audio
dataset = load_dataset("bofenghuang/asr-dummy", "cs", split="test")
sample, text = dataset[0]["audio"], dataset[0]["text"]

# Ground truth text
print(text)
# Aber sei ihnen nicht böse, Habibi, vergib ihnen, sie vergaßen die Liebe, sie vergaßen die Bibel, 
# wünsch ihnen den Frieden. Nous allons construire des radiotélescopes géants comme celui-ci, 
# qui est mon préféré. Questa è un'immagine di Cairo Open City, una mostra che il museo Folkwang di 
# Essen ha dedicato al ruolo della mobile photography nella primavera Araba.

# Extract feautres
input_features = processor(
    sample["array"], sampling_rate=sample["sampling_rate"], return_tensors="pt"
).input_features


# Generate tokens
predicted_ids = model.generate(
    input_features.to(device, dtype=torch_dtype),
    max_new_tokens=128,
)

# Detokenize to text
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
print(transcription)
#  Aber sei ihnen nicht böse, Habibi, vergib ihn. Sie vergaßen die Liebe, sie vergaßen die Liebe. 
# Wünsche ihnen dem Frieden. Nous allons construire des radiotelescopes géants, comme celui-ci qui 
# est mon préféré. Esta es una imagen de Cairo Open City, una muestra que el Museo Folk Punk de Essen 
# ha dedicado al ruolo de la mobile fotografía en la primavera árabe.

# Dive in generated tokens
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=False)[0]
print(transcription)
# <|de|> Aber sei ihnen nicht böse, Habibi, vergib ihn. Sie vergaßen die Liebe, sie vergaßen die Liebe. 
# Wünsche ihnen dem Frieden.<|fr|> Nous allons construire des radiotelescopes géants, comme celui-ci qui 
# est mon préféré.<|es|> Esta es una imagen de Cairo Open City, una muestra que el Museo Folk Punk de Essen 
# ha dedicado al ruolo de la mobile fotografía en la primavera árabe.

Evaluation

English

French

Spanish

German