startelelogic/whisper-medium-ccaas

Whisper-Medium (Fine-Tuned on Svarah Dataset)

📘 Model Overview

This repository contains a fine-tuned and fully merged version of OpenAI’s Whisper-Medium automatic speech recognition (ASR) model.

The model was fine-tuned on the Svarah dataset, which consists of Indian-accented contact center call recordings and their corresponding transcriptions.
Fine-tuning was performed using LoRA (Low-Rank Adaptation) in PyTorch and later merged into the base Whisper-Medium weights, resulting in a single, standalone checkpoint.

✅ No LoRA adapters required at inference
✅ Fully offline usage
✅ No OpenAI API dependency
✅ Optimized for Indian English contact center speech
✅ Compatible with PyTorch, Transformers, and CTranslate2 (after conversion)

---

🧠 Model Description

• Base model: openai/whisper-medium
• Architecture: Encoder–Decoder Transformer (Whisper)
• Parameters: ~769M
• Task: Automatic Speech Recognition (ASR)
• Input: 16 kHz mono audio
• Output: Text transcription
• Primary Accent: Indian English
• Domain: Contact center / customer support calls
• Languages: English (primary), multilingual capability retained

This model improves transcription accuracy for Indian-accented conversational speech, particularly in telephony and contact center environments, while retaining Whisper’s general robustness.

---

🎧 Training Data

Svarah Dataset

This model was fine-tuned on the Svarah dataset, which includes:

• Real-world contact center call recordings
• Multiple Indian accents and dialects
• Conversational and spontaneous speech
• Human-verified transcription quality
• Low-bandwidth, telephony-style audio characteristics
• The dataset focuses on realistic customer support interactions rather than studio-quality speech.

🔄 Data Processing

During fine-tuning, the following preprocessing steps were applied:

• Audio normalized and resampled to 16 kHz mono
• Text cleaned and standardized through normalization
• Long call recordings segmented into Whisper-compatible chunks
• Overlapping windows added to preserve contextual continuity
• Noise and silence trimming applied where required

---

Training Details

Preprocessing Pipeline

Before training, all audio and transcripts were standardized:

Audio

• Converted to 16 kHz mono
• Normalized and peak-limited
• Split into Whisper-compatible segments
• Applied overlapping chunk windows for context retention
• Removed long silences and excessive noise

Text

• Lowercased and normalized
• Removed filler markers where applicable
• Standardized spellings for Indian English
• Cleaned punctuation, spacing, and verbal artifacts
• Mapped special characters to ASCII equivalents

Fine-Tuning Method

• Parameter-Efficient Fine-Tuning using LoRA
• Base Whisper-Medium weights were frozen
• LoRA applied to attention projection layers:
  • q_proj, k_proj, v_proj, o_proj

LoRA Configuration

• Rank: 16
• Alpha: 32
• Dropout: 0.05
• Target Modules: Q/K/V projection layers in the Whisper transformer This allowed the model to learn accent-, vocabulary-, and domain-specific characteristics without modifying the full 769M parameter base.

Training Frameworks

• Framework: PyTorch
• Library: Hugging Face Transformers + PEFT (LoRA)
• Hardware: Multi-GPU training (NVIDIA A100 / RTX 4090)
• Precision: Mixed-precision (fp16) The combination of LoRA and mixed-precision training ensured efficient GPU utilization and stable training on large audio datasets.

---

Model Merging

After training:

• LoRA adapters were merged into the base model
• Adapters were removed using merge_and_unload
• Final model saved as a single unified Whisper checkpoint

The final exported model is a standalone checkpoint, identical in structure to Whisper-Medium but with improved recognition performance on Indian accents.

📊 Evaluation Metrics

This model was evaluated using industry-standard transcription metrics focused on accuracy, stability, and performance in real-world telephony environments. The primary metrics used are Word Error Rate (WER) and Character Error Rate (CER).

1. Word Error Rate (WER)

   Performance Summary

   Dataset Split	WER (Whisper-Medium Base)	WER (Fine-Tuned Model)
   Svarah Test Set	21.3%	13.4%
   Hinglish Subset	26.8%	15.9%
   Regional English (Gujarati, Marathi, Telugu)	27.4%	17.2%
   Noisy Telephony Audio	24.6%	14.1%

   Improvement: The fine-tuned model shows an average 34–42% relative WER reduction, especially on code-switched (Hinglish) and noisy contact center audio.

2. Character Error Rate (CER)

   Performance Summary

   Dataset Split	CER (Base)	CER (Fine-Tuned)
   Svarah Test Set	13.7%	8.2%
   Hinglish Subset	17.4%	9.9%
   Regional Accents	18.2%	11.1%
   Noisy Telephony Audio	16.1%	9.0%

   The improvements in CER confirm that the fine-tuned model handles accented pronunciation, speech rate variation, and irregular spacing more effectively.

3. Qualitative Improvements

   Accent Handling

   • Better recognition of Hindi, Gujarati, Marathi, Telugu, Tamil, and Bengali accents
   • More stable decoding of Indian English pronunciation patterns
   • Reduced errors with long or complex words Code-Switching Performance
   • Significant improvement in Hinglish transcription accuracy
   • Handles fast switching between languages with fewer substitutions Noise Robustness
   • Improved performance with low-bitrate telephony audio
   • Fewer hallucinations during background noise or overlaps
   • Better segmentation and continuity in long conversations

4. Evaluation Methodology

   Evaluation was performed using:

   • The Svarah dataset test split
   • Additional manually curated Hinglish test samples
   • Noisy, real-world telephony recordings
   • Standard Hugging Face WER/CER evaluation scripts Transcriptions from the base Whisper-Medium model were compared directly against the fine-tuned model to measure relative performance gain.

Inference Usage

PyTorch / Transformers

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import torch

model_name = "startelelogic/whisper-medium-ccaas"
processor = WhisperProcessor.from_pretrained(model_name)
model = WhisperForConditionalGeneration.from_pretrained(model_name).to("cuda")

inputs = processor("audio.wav", return_tensors="pt", sampling_rate=16000)
with torch.no_grad():
    generated_ids = model.generate(inputs.input_features.to("cuda"))

text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(text)

Citation

@misc{radford2022whisper, doi = {10.48550/ARXIV.2212.04356}, url = {https://arxiv.org/abs/2212.04356}, author = {Radford, Alec and Kim, Jong Wook and Xu, Tao and Brockman, Greg and McLeavey, Christine and Sutskever, Ilya}, title = {Robust Speech Recognition via Large-Scale Weak Supervision}, publisher = {arXiv}, year = {2022}, copyright = {arXiv.org perpetual, non-exclusive license} }