Speech recognition using OpenAI’s Whisper model

Speech recognition is a field in artificial intelligence that allows computers to understand human speech and convert it into text. The technology is used in devices such as Alexa and various chatbot applications. The most common thing we do is voice transcription, which can be converted into transcripts or subtitles.

#Recent developments in state-of-the-art models such as wav2vec2, Conformer, and Hubert have significantly advanced the field of speech recognition. These models employ techniques that learn from raw audio without manually labeling the data, allowing them to efficiently use large datasets of unlabeled speech. They have also been extended to use up to 1,000,000 hours of training data, well beyond the traditional 1,000 hours used in academic supervised datasets, but models pretrained in a supervised manner across multiple datasets and domains have been found to perform better Robustness and generalization to held datasets, so performing tasks such as speech recognition still requires fine-tuning, which limits their full potential. To solve this problem OpenAI developed Whisper, a model that utilizes weak supervision methods.

This article will explain the types of data sets used for training, the training methods of the model, and how to use Whisper

Whisper model introduction

Using the data set:

The Whisper model was trained on a dataset of 680,000 hours of labeled audio data, including 117,000 hours of speech in 96 different languages ​​and 125,000 hours of translation data from "any language" to English. The model leverages Internet-generated text that is generated by other automatic speech recognition systems (ASR) rather than created by humans. The dataset also includes a language detector trained on VoxLingua107, a collection of short speech clips extracted from YouTube videos and tagged based on the language of the video title and description, with additional steps to remove false positives .

Model:

The main structure used is the encoder-decoder structure.

Resampling: 16000 Hz

Feature extraction method: Calculate an 80-channel log Mel spectrogram representation using a 25 ms window and a 10 ms step.

Feature normalization: The input is globally scaled to between -1 and 1 and has an approximately zero mean on the pre-trained dataset.

Encoder/Decoder: The encoder and decoder of this model use Transformers.

Procedure of the encoder:

The encoder first processes the input representation using a stem containing two convolutional layers (filter width 3), using the GELU activation function.

The stride of the second convolutional layer is 2.

The sinusoidal position embedding is then added to the output of the stem and the encoder Transformer block is applied.

Transformers use pre-activated residual blocks, and the output of the encoder is normalized using a normalization layer.

Model block diagram:

Decoding process:

In the decoder, learning position embedding and binding input and output markers are used express.

The encoder and decoder have the same width and number of Transformers blocks.

Training

To improve the scaling properties of the model, it is trained on different input sizes.

Train the model through FP16, dynamic loss scaling, and data parallelism.

Using AdamW and gradient norm clipping, the linear learning rate decays to zero after warming up the first 2048 updates.

Use a batch size of 256 and train the model for 220 updates, which is equivalent to two to three forward passes on the dataset.

Since the model was trained for only a few epochs, overfitting was not a significant issue, and no data augmentation or regularization techniques were used. This instead relies on diversity within large datasets to promote generalization and robustness.

Whisper has demonstrated good accuracy on previously used datasets and has been tested against other state-of-the-art models.

Advantages:

• Whisper has been trained on real data as well as data used on other models and with weak supervision.
• The accuracy of the model was tested against human listeners and its performance evaluated.
• It detects unvoiced areas and applies NLP technology to correctly punctuate the transcript.
• The model is scalable and allows extracting transcripts from audio signals without dividing the video into chunks or batches, thus reducing the risk of missing sounds.
• The model achieves higher accuracy on various data sets.

Comparative results of Whisper on different data sets, compared with wav2vec, it has achieved the lowest word error rate so far

The model was not tested on the timit dataset, so in order to check its word error rate, we will demonstrate here how to use Whisper to self-verify the timit dataset, that is, use Whisper to build our own speech recognition application .

Using Whisper Model for Speech Recognition

The TIMIT Reading Speech Corpus is a collection of speech data specifically used for acoustic speech research and the development and evaluation of automatic speech recognition systems. It includes recordings of 630 speakers from the eight major dialects of American English, each reading ten phonetically rich sentences. The corpus includes time-aligned orthographic, phonetic, and word transcriptions as well as 16-bit, 16kHz speech waveform files for each voice. The corpus was developed by the Massachusetts Institute of Technology (MIT), SRI International (SRI), and Texas Instruments (TI). TIMIT corpus transcriptions have been manually verified, with testing and training subsets specified to balance phonetic and dialect coverage.

Installation:

!pip install git+https://github.com/openai/whisper.git
 !pip install jiwer
 !pip install datasets==1.18.3

The first command will install all the dependencies required by the whisper model. jiwer is used to download the text error rate package. The datasets are provided by hugface. You can download the timit dataset.

Import library

import whisper
 from pytube import YouTube
 from glob import glob
 import os
 import pandas as pd
 from tqdm.notebook import tqdm

Load timit data set

from datasets import load_dataset, load_metric
 
 timit = load_dataset("timit_asr")

Calculate the Word error rate under different model sizes

Consider filtering English data and non-English data To meet the needs, we choose to use a multi-language model here instead of a model specifically designed for English.

But the TIMIT data set is pure English, so we have to apply the same language detection and recognition process. In addition, the TIMIT data set has been divided into training and verification sets, and we can use it directly.

To use Whisper, we must first understand the parameters, size and speed of different models.

Load model

model = whisper.load_model('tiny')

tiny can be replaced with the model name mentioned above.

Function to define language detector

def lan_detector(audio_file):
 print('reading the audio file')
 audio = whisper.load_audio(audio_file)
 audio = whisper.pad_or_trim(audio)
 mel = whisper.log_mel_spectrogram(audio).to(model.device)
 _, probs = model.detect_language(mel)
 if max(probs, key=probs.get) == 'en':
 return True
 return False

Function to convert speech to text

def speech2text(audio_file):
 text = model.transcribe(audio_file)
 return text["text"]

Run the above function under different model sizes, timit the word errors obtained by training and testing The rate is as follows:

Transcribe speech from u2b

Compared with other speech recognition models, Whisper can not only recognize speech, but also interpret the content of a person’s speech punctuation intonation, and insert appropriate punctuation marks. We will use u2b’s video for testing below.

Here we need a package pytube, which can easily help us download and extract audio

def youtube_audio(link):
 youtube_1 = YouTube(link)
 videos = youtube_1.streams.filter(only_audio=True)
 
 name = str(link.split('=')[-1])
 out_file = videos[0].download(name)

 link = name.split('=')[-1]
 new_filename = link+".wav"
 print(new_filename)
 os.rename(out_file, new_filename)
 print(name)
 return new_filename,link

After obtaining the wav file, we can apply the above function to extract text from it.

Summary

The code of this article is here

https://drive.google.com/file/d/1FejhGseX_S1Ig_Y5nIPn1OcHN8DLFGIO/view

There are many more The operation can be completed with Whisper, and you can try it yourself based on the code in this article.

The above is the detailed content of Speech recognition using OpenAI’s Whisper model. For more information, please follow other related articles on the PHP Chinese website!