Oriserve / Whisper-Hindi2Hinglish-Prime

Whisper-Hindi2Hinglish-Prime:

Table of Contents:

• Key Features
• Training
  • Data
  • Finetuning
• Usage
• Performance Overview
  • Qualitative Performance Overview
  • Quantitative Performance Overview
• Miscellaneous

Key Features:

1. Hinglish as a language : Added ability to transcribe audio into spoken Hinglish language reducing chances of grammatical errors
2. Whisper Architecture : Based on the whisper architecture making it easy to use with the transformers package
3. Better Noise handling : The model is resistant to noise and thus does not return transcriptions for audios with just noise
4. Hallucination Mitigation : Minimizes transcription hallucinations to enhance accuracy.
5. Performance Increase : ~39% average performance increase versus pretrained model across benchmarking datasets

Training:

Data:

• Duration : A total of ~550 Hrs of noisy Indian-accented Hindi data was used to finetune the model.
• Collection : Due to a lack of ASR-ready hinglish datasets available, a specially curated proprietary dataset was used.
• Labelling : This data was then labeled using a SOTA model and the transcriptions were improved by human intervention.
• Quality : Emphasis was placed on collecting noisy data for the task as the intended use case of the model is in Indian environments where background noise is abundant.
• Processing : It was ensured that the audios are all chunked into chunks of length <30s, and there are at max 2 speakers in a clip. No further processing steps were done so as to not change the quality of the source data.

Finetuning:

• Novel Trainer Architecture : A custom trainer was written to ensure efficient supervised finetuning, with custom callbacks to enable higher observability during the training process.
• Custom Dynamic Layer Freezing : Most active layers were identified in the model by running inference on a subset of the training data using the pre-trained models. These layers were then kept frozen during the training process while all the other layers were kept frozen. This enabled faster convergence and efficient finetuning
• Deepspeed Integration : Deepspeed was also utilized to speed up, and optimize the training process.

Performance Overview

Qualitative Performance Overview

Quantitative Performance Overview

Note :

• The below WER scores are for Hinglish text generated by our model and the original whisper model
• To check our model's real-world performance against other SOTA models please head to our Speech-To-Text Arena arena space.

Usage:

Using Transformers

• To run the model, first install the Transformers library

pip install -U transformers

• 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

# Set device (GPU if available, otherwise CPU) and precision
device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

# Specify the pre-trained model ID
model_id = "Oriserve/Whisper-Hindi2Hinglish-Prime"

# Load the speech-to-text model with specified configurations
model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, 
    torch_dtype=torch_dtype,        # Use appropriate precision (float16 for GPU, float32 for CPU)
    low_cpu_mem_usage=True,         # Optimize memory usage during loading
    use_safetensors=True            # Use safetensors format for better security
)
model.to(device)                    # Move model to specified device

# Load the processor for audio preprocessing and tokenization
processor = AutoProcessor.from_pretrained(model_id)

# Create speech recognition pipeline
pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    torch_dtype=torch_dtype,
    device=device,
    generate_kwargs={
        "task": "transcribe",       # Set task to transcription
        "language": "en"            # Specify English language
    }
)

# Process audio file and print transcription
sample = "sample.wav"               # Input audio file path
result = pipe(sample)               # Run inference
print(result["text"])               # Print transcribed text

Using Flash Attention 2

Flash-Attention 2 can be used to make the transcription fast. If your GPU supports Flash-Attention you can use it by, first installing Flash Attention:

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

• Once installed you can then load the model using the below code:

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

Using the OpenAI Whisper module

• First, install the openai-whisper library

pip install -U openai-whisper tqdm

• Convert the huggingface checkpoint to a pytorch model

import torch
from transformers import AutoModelForSpeechSeq2Seq
import re
from tqdm import tqdm
from collections import OrderedDict
import json

# Load parameter name mapping from HF to OpenAI format
with open('convert_hf2openai.json', 'r') as f:
    reverse_translation = json.load(f)

reverse_translation = OrderedDict(reverse_translation)

def save_model(model, save_path):
    def reverse_translate(current_param):
        # Convert parameter names using regex patterns
        for pattern, repl in reverse_translation.items():
            if re.match(pattern, current_param):
                return re.sub(pattern, repl, current_param)

    # Extract model dimensions from config
    config = model.config
    model_dims = {
        "n_mels": config.num_mel_bins,           # Number of mel spectrogram bins
        "n_vocab": config.vocab_size,            # Vocabulary size
        "n_audio_ctx": config.max_source_positions,    # Max audio context length
        "n_audio_state": config.d_model,         # Audio encoder state dimension
        "n_audio_head": config.encoder_attention_heads,  # Audio encoder attention heads
        "n_audio_layer": config.encoder_layers,   # Number of audio encoder layers
        "n_text_ctx": config.max_target_positions,     # Max text context length
        "n_text_state": config.d_model,          # Text decoder state dimension
        "n_text_head": config.decoder_attention_heads,  # Text decoder attention heads
        "n_text_layer": config.decoder_layers,    # Number of text decoder layers
    }

    # Convert model state dict to Whisper format
    original_model_state_dict = model.state_dict()
    new_state_dict = {}

    for key, value in tqdm(original_model_state_dict.items()):
        key = key.replace("model.", "")          # Remove 'model.' prefix
        new_key = reverse_translate(key)         # Convert parameter names
        if new_key is not None:
            new_state_dict[new_key] = value

    # Create final model dictionary
    pytorch_model = {"dims": model_dims, "model_state_dict": new_state_dict}

    # Save converted model
    torch.save(pytorch_model, save_path)

# Load Hugging Face model
model_id = "Oriserve/Whisper-Hindi2Hinglish-Prime"
model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, 
    low_cpu_mem_usage=True,        # Optimize memory usage
    use_safetensors=True           # Use safetensors format
)

# Convert and save model
model_save_path = "Whisper-Hindi2Hinglish-Prime.pt"
save_model(model,model_save_path)

• Transcribe

import whisper
# Load converted model with Whisper and transcribe
model = whisper.load_model("Whisper-Hindi2Hinglish-Prime.pt")
result = model.transcribe("sample.wav")
print(result["text"])

Miscellaneous

This model is from a family of transformers-based ASR models trained by Oriserve. To compare this model against other models from the same family or other SOTA models please head to our Speech-To-Text Arena . To learn more about our other models, and other queries regarding AI voice agents you can reach out to us at our email [email protected]