🎙️ CoSHE-Eval: A Code-Switching ASR Benchmark for Hindi–English Speech

🧠 Overview

CoSHE-Eval is an evaluation dataset curated for testing Automatic Speech Recognition (ASR) systems on Hindi-English code-mixed speech.It focuses on bilingual conversational contexts commonly found in India, where Hindi (in Devanagari) and English (in Latin script) co-occur naturally within the same utterance.

Detailed Blog: CoSHE-Eval Blog

---

Technical Specifications

Attribute	Description
Total Samples	1985
Total Duration	~30 hours
Minimum Segment Length	0.60 seconds
Maximum Segment Length	59.8 seconds
Mean Segment Length	53.3 seconds
Median Segment Length	56.9 seconds
Timestamp Validation	Incremental and aligned with audio duration
Speaker Segmentation	Maintains full utterances; no mid-sentence cuts

---

📂 Dataset Structure

Column	Description
audio_file_name	Unique name or ID of the audio sample
transcription	Verified ground-truth transcription (Hindi-English code-mixed)
audio	The corresponding audio waveform

All audio files are provided in .wav format and perfectly aligned with their corresponding transcriptions.

---

⚙️ Example: Computing Word Error Rate (WER)

Below is an example comparing a ground truth transcript with a test model transcript to compute the Word Error Rate (WER):

import evaluate

# Ground truth vs test model sentences
reference_text = "आज मैंने new laptop खरीदा और performance बहुत अच्छी है"
predicted_text = "आज मैंने new laptop लिया और performance अच्छी है"

print("Ground Truth Transcript:\n", reference_text)
print("\nTest Model Transcript:\n", predicted_text)

# Compute WER
wer_metric = evaluate.load("wer")
wer_score = wer_metric.compute(predictions=[predicted_text], references=[reference_text])
print(f"\nWord Error Rate (WER): {wer_score:.3f}")

Output:

Ground Truth Transcript:
आज मैंने new laptop खरीदा और performance बहुत अच्छी है

Test Model Transcript:
आज मैंने new laptop लिया और performance अच्छी है

Word Error Rate (WER): 0.2

This demonstrates how the Ground Truth 10 dataset can be used to quantitatively assess ASR model accuracy using standard evaluation metrics such as WER.

🚀 Usage Example You can load the dataset directly using the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("soketlabs/CoSHE-Eval", split="eval")
print(dataset[0])

🧠 About Soket AI

Soket AI is a deep-tech AI research and innovation company committed to advancing sovereign, ethical, and inclusive artificial intelligence.Our mission is to build cutting-edge AI systems that empower industries, researchers, and citizens alike — spanning domains such as speech recognition, defense, healthcare, education, and Indic language intelligence.

At Soket AI, we believe in AI made for people, by people, fostering trust, transparency, and accessibility at every layer.

Learn more: https://soket.ai/

🏛️ About Project EKΛ

Project EKΛ (pronounced Eka, meaning “One” in Sanskrit) represents India’s bold leap toward sovereign and inclusive intelligence. It is an initiative to build foundational AI systems that speak every language, reflect every culture, and empower every citizen. Rooted in India’s linguistic and cultural diversity, EKΛ is designed to advance AI that is both globally competitive and locally grounded.

Project EKΛ is being developed in collaboration with IIT Gandhinagar, bringing together cutting-edge academic research and real-world AI engineering. This collaboration ensures strong scientific foundations, rigorous evaluation, and responsible development aligned with India’s multilingual and societal needs.

At the core of EKΛ lies a 120-billion-parameter multilingual foundation model—a state-of-the-art large language model (LLM) engineered to understand and generate content across all major Indic languages, English, and their code-mixed variants. Built in India, for the world, EKΛ aims to set a new global benchmark for humane, inclusive, and culturally grounded AI systems.

🔗 Join the initiative: https://eka.soket.ai/

💬 Contact

For any queries, collaborations, or feedback related to this dataset, please reach out via:

📧 Email: connect@soket.ai