ShimogaAIteam/conversational_kannada_stt

Conversational Kannada STT

This dataset contains corrected transcriptions of conversational Kannada speech,
prepared specifically for fine-tuning Whisper models on conversational and dialectal Kannada.

Unlike many ASR datasets, this release provides pre-computed Whisper input features (log-Mel spectrograms)
so you can train/fine-tune Whisper models without raw audio processing.

---

Dataset Creation

• Source Audio: Publicly available YouTube videos in Kannada.
• Initial Transcripts: Extracted using YouTube’s auto-transcription system.
• Corrections:
  • First-pass corrections with Gemini Pro for common errors.
  • Final manual corrections by ShimogaAIteam, focusing on dialects & colloquial usage.

This process ensures the dataset is closer to real-world conversational Kannada.

---

Dataset Details

• Language: Kannada (kn)
• Domain: Conversational & Dialectal Speech
• Rows: ~5K
• Audio: ❌ Not included
• Features: ✅ Pre-computed Whisper log-Mel spectrograms

Columns

Column	Description
sentence	Corrected transcription in Kannada
input_features	Pre-computed Whisper log-Mel spectrogram features
labels	Tokenized labels for Whisper training

---

Intended Use

• Fine-tuning Whisper for Kannada conversational ASR
• Coverage of dialects and colloquial Kannada
• Research on low-resource ASR in Indian languages

---

Usage Example

from datasets import load_dataset

dataset = load_dataset("ShimogaAIteam/conversational_kannada_stt")

print(dataset)
print(dataset["train"][0]["sentence"])
Citation
If you use this dataset, please cite:

bibtex
Copy code
@dataset{shimogaai_conversational_kannada_stt,
  title        = {Conversational Kannada STT (Corrected)},
  author       = {ShimogaAIteam},
  year         = {2025},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/ShimogaAIteam/conversational_kannada_stt}
}
Acknowledgements
ShimogaAIteam for manual corrections & dataset preparation

YouTube for providing source audio & transcripts

Gemini Pro for transcript refinement assistance

The open-source Whisper community for enabling Kannada ASR research