IndiVoice-DeepASR: Indian-Accented Speech Recognition

Bridging the Accent Gap in Modern ASR with Whisper + LoRA

Note

Active Repository Notice: The primary development, active issues, and latest updates for this project are maintained at the official organization repository: PxA-Labs/IndiVoice-DeepASR. Please direct all issues, feature requests, and contributions there.

Explore the Code • Launch Colab • Launch Kaggle

Important

Resilience Update (v1.8): Implemented high-frequency checkpointing (every 100 steps) and auto-resumption to protect training progress against Colab or Kaggle runtime disconnections. Resolved load_best_model_at_end compatibility issues.

---

Overview

Current commercial automatic speech recognition (ASR) systems demonstrate a 20-30% performance degradation when processing Indian English accents. IndiVoice-DeepASR is a research-driven project that fine-tunes OpenAI's Whisper models using Low-Rank Adaptation (LoRA) to achieve state-of-the-art accuracy across diverse Indian linguistic profiles.

Key Features

• Fault-Tolerant Training: Automatic checkpoint detection and seamless resumption to safeguard progress during remote GPU training disconnections.
• Parameter Efficiency: Fine-tune with less than 2% of total parameters using Parameter-Efficient Fine-Tuning (PEFT) techniques.
• Accent Localization: Optimized to handle Hindi, Tamil, Kannada, Bengali, and Punjabi regional accents.
• Robust Audio Pipeline: Multi-layered AudioDecoder logic for stable preprocessing across diverse computing environments.
• Enhanced Accuracy: Accomplishes significant Word Error Rate (WER) reductions compared to the base Whisper model.

---

Technical Stack and Architecture

Model Backbone	Optimization	Audio Engine
Cloud Compute	Deployment	Infrastructure

---

Pipeline Flow

graph LR
    A[Raw Audio] --> B(Standardization: 16kHz Mono)
    B --> C{IndiVoice Engine}
    C --> D[Whisper Backbone]
    C --> E[LoRA Adapters]
    D & E --> F[Optimized Transcripts]
    F --> G[Metric Analysis: WER/CER]

Loading

---

Getting Started

1. Cloud-Based Training (Recommended)

Choose your preferred platform for free GPU access:

• Colab Gateway: Best for initial setup and rapid experimentation.
• Kaggle Runner: Optimized for long-running training. Includes setup_kaggle.sh for environment configuration.

2. Local Setup and Execution

# Clone the repository
git clone https://github.com/PxA-Labs/IndiVoice-DeepASR.git
cd IndiVoice-DeepASR

# Install dependencies
pip install -r requirements.txt

# Preprocess the Svarah dataset
python src/preprocess.py --hf_dataset ai4bharat/Svarah --output_dir data/processed

# Train the model (Auto-resumes from latest checkpoint if present)
python src/train.py --output_dir models/indian-accent-lora

---

Repository Structure

IndiVoice-DeepASR/
├── assets/            # Branding and visual elements
├── kaggle/            # Kaggle training utilities and scripts
├── src/               # Core codebase for training, preprocessing, and deployment
├── notebooks/         # Jupyter Notebooks for experimentation
├── data/              # Dataset symlinks and manifests
├── models/            # Checkpoints and serialized weights
└── paper/             # Source files for academic publication

---

Academic Citation

If you use this work in your research, please cite:

@misc{indivoice2026,
  author = {Purvansh Joshi and Archit Mittal},
  title = {IndiVoice-DeepASR: Efficient Adaptation of Multilingual Speech Models for Indian Accents},
  year = {2026},
  publisher = {GitHub},
  howpublished = {\url{https://github.com/PxA-Labs/IndiVoice-DeepASR}}
}

---

Built for the Indian Speech Recognition Research Community