deepdml/iisc-mile-tamil-asr

IISc-MILE Tamil ASR Corpus

Summary

The IISc-MILE Tamil ASR Corpus is a large-scale transcribed speech dataset designed for training Automatic Speech Recognition (ASR) systems for the Tamil language. It contains approximately 150 hours of high-quality read speech collected from 531 speakers in a noise-free recording environment using professional USB microphones.

This corpus was published by the Medical Intelligence and Language Engineering (MILE) Lab at the Indian Institute of Science (IISc), Bangalore, India.

---

Dataset Structure

Data Splits

Split	Utterances	Approx. hours
train	77,314	~131 h
test	12,087	~20 h
Total	89,401	~150 h

Features

Feature	Type	Description
id	string	Unique utterance identifier (base filename)
audio	Audio	WAV audio at 16 kHz, 16-bit, mono, PCM
transcription	string	Corresponding Tamil text in UTF-8 Unicode
duration	float32	Duration of the audio clip in seconds

Data Fields

• id: The base filename shared by the .wav and .txt files (e.g., spk001_utt042).
• audio: Audio object containing the waveform array, sampling rate (16000 Hz), and original file path.
• transcription: The Tamil text transcription of the audio recording, encoded in UTF-8 Unicode.
• duration: Duration of the audio recording in seconds, computed from the audio file header.

---

Dataset Creation

Source Data

Speech was recorded in a controlled, noise-free studio environment using high-quality USB microphones. Speakers read pre-selected text passages in Tamil. All recordings are in WAV format (16 kHz, 16-bit, mono, PCM).

Annotations

Transcriptions are provided as plain text files in UTF-8 Unicode encoding, one file per audio recording. Each transcript corresponds exactly to its paired audio file by filename.

Recording Conditions

Property	Value
Format	WAV (PCM)
Sample Rate	16,000 Hz
Bit Depth	16-bit
Channels	Mono
Environment	Noise-free studio
Microphone	High-quality USB

Speaker Statistics

Property	Value
Total speakers	531
Total duration	~150 hours
Speech type	Read speech

---

Usage

Load the Dataset

from datasets import load_dataset

# Load full dataset
dataset = load_dataset("your-username/mile-tamil-asr")

# Load a specific split
train_data = load_dataset("your-username/mile-tamil-asr", split="train")
test_data  = load_dataset("your-username/mile-tamil-asr", split="test")

Inspect a Sample

sample = train_data[0]

print(sample["id"])             # e.g. "spk001_utt001"
print(sample["transcription"])  # Tamil text
print(sample["duration"])     # e.g. 4.352

# Play audio (in a notebook)
import IPython.display as ipd
ipd.Audio(sample["audio"]["array"], rate=sample["audio"]["sampling_rate"])

Filter by Duration

# Keep only utterances between 2 and 15 seconds
filtered = train_data.filter(lambda x: 2.0 <= x["duration"] <= 15.0)

---

Considerations for Use

Social Impact

This dataset contributes to advancing ASR technology for Tamil, a classical Dravidian language spoken by over 80 million people worldwide. Improved ASR for Tamil supports accessibility tools, voice interfaces, and language technology for a historically underserved language community.

Limitations

• Speech is read speech only; spontaneous or conversational speech patterns are not represented.
• Recordings were made in controlled studio conditions, so models trained on this corpus may be less robust to real-world noise.
• Speaker demographics (age, gender distribution) are not detailed in the original release.

---

Citation

If you use this dataset, please cite the following works:

@misc{mile_1,
  doi       = {10.48550/ARXIV.2207.13331},
  url       = {https://arxiv.org/abs/2207.13331},
  author    = {A, Madhavaraj and Pilar, Bharathi and G, Ramakrishnan A},
  title     = {Subword Dictionary Learning and Segmentation Techniques for
               Automatic Speech Recognition in Tamil and Kannada},
  publisher = {arXiv},
  year      = {2022},
}

@misc{mile_2,
  doi       = {10.48550/ARXIV.2207.13333},
  url       = {https://arxiv.org/abs/2207.13333},
  author    = {A, Madhavaraj and Pilar, Bharathi and G, Ramakrishnan A},
  title     = {Knowledge-driven Subword Grammar Modeling for
               Automatic Speech Recognition in Tamil and Kannada},
  publisher = {arXiv},
  year      = {2022},
}

---

Dataset Card Authors

This dataset card was prepared to accompany the HuggingFace upload of the IISc-MILE Tamil ASR Corpus (OpenSLR SLR127). The original corpus was created and published by the MILE Lab, Indian Institute of Science, Bangalore.