Transducens/error-preserving-whisper-distilled

Whisper Model for Incorrect English Phrases

Overview

This fine-tuned version of OpenAI’s Whisper model is specifically trained to handle incorrect English phrases. It is designed to transcribe and process non-standard or erroneous English input, including mispronunciations, grammatical mistakes, slang, and non-native speaker errors. This model helps improve transcription accuracy in scenarios where speakers use incorrect or informal English, making it useful in language learning, transcription of casual conversations, or analyzing spoken communication from non-native English speakers.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 28
• eval_batch_size: 28
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 50
• training_steps: 100000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.5189	0.4444	500	1.1913	25.9108
1.1727	0.8889	1000	0.9531	24.5396
1.1341	1.3333	1500	0.8688	22.2761
1.0152	1.7778	2000	0.8174	20.8792
1.0589	2.2222	2500	0.7855	20.7595
0.9793	2.6667	3000	0.7611	22.2846
0.9594	3.1111	3500	0.7442	20.3860
1.0031	3.5556	4000	0.7303	18.5045
0.9525	4.0	4500	0.7199	18.1054
0.8729	4.4444	5000	0.7105	19.3170
1.0031	4.8889	5500	0.7028	19.7446
0.9273	5.3333	6000	0.6966	19.7189
0.9174	5.7778	6500	0.6896	18.4475
0.8842	6.2222	7000	0.6839	18.4361

Usage Guide

This project was executed on an Ubuntu 22.04.3 system running Linux kernel 6.8.0-40-generic.

Whisper large-v3 is supported in Hugging Face Transformers. To run the model, first install the Transformers library. For this example, we'll also install Hugging Face Datasets to load toy audio dataset from the Hugging Face Hub, and Hugging Face Accelerate to reduce the model loading time:

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:

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

def download_adapter_model():
  model_name = "whisper-v3-LoRA-en_students"
  print(f"Downloading the adapter model '{model_name}' from the Hugging Face Hub.", flush=True)

  # Define the path for the directory
  local_directory = os.path.expanduser("~/.cache/huggingface/hub")

  # Check if the directory exists
  if not os.path.exists(local_directory):
    # If it doesn't exist, create it
      os.makedirs(local_directory)
      print(f"Directory '{local_directory}' created.", flush=True)
  else:
    print(f"Directory '{local_directory}' already exists.", flush=True)

  repo_id = f"Transducens/{model_name}"
  repo_adapter_dir = f"{model_name}/checkpoint-5000/adapter_model"
  repo_filename_config = f"{repo_adapter_dir}/adapter_config.json"
  repo_filename_tensors = f"{repo_adapter_dir}/adapter_model.safetensors"

  adapter_config = hf_hub_download(repo_id=repo_id, filename=repo_filename_config, local_dir=local_directory)
  adapter_model_tensors = hf_hub_download(repo_id=repo_id, filename=repo_filename_tensors, local_dir=local_directory)

  print(f"Dowloaded the adapter model '{model_name}' from the Hugging Face Hub.", flush=True)

  return os.path.join(local_directory, repo_adapter_dir)

peft_model_id = adapter_path # Use the same model ID as before.
peft_config = PeftConfig.from_pretrained(peft_model_id)
model = WhisperForConditionalGeneration.from_pretrained(
peft_config.base_model_name_or_path, load_in_8bit=False)

model = PeftModel.from_pretrained(model, peft_model_id)
model.generation_config.language = "<|en|>"
model.generation_config.task = "transcribe"

tokenizer = WhisperTokenizer.from_pretrained("openai/whisper-large-v3", task="transcribe")
feature_extractor = WhisperFeatureExtractor.from_pretrained("openai/whisper-large-v3")

pipe = pipeline(model=model, tokenizer=tokenizer, feature_extractor=feature_extractor, task="automatic-speech-recognition", device=device)

Framework versions

• PEFT 0.11.1
• Transformers 4.42.4
• Pytorch 2.1.0+cu118
• Datasets 2.20.0
• Tokenizers 0.19.1