Ranjan-Hindi33min

Owner: @BBSRguy
Created: 2025-06-03
Year: 2025
Language: Hindi 🇮🇳
Region Focus: Odisha, India
Sample Rate Variants: 16 kHz, 24 kHz, 32 kHz
Total Files: 29 pairs (speech + text)
Duration: Approximately 33 minutes of speech

📜 Description

Ranjan-Hindi33min is a meticulously curated dataset comprising high-quality Hindi speech samples and their corresponding textual transcriptions. This dataset is designed to support various speech processing tasks, including Automatic Speech Recognition (ASR), Text-to-Speech (TTS) synthesis, and speech-text alignment, with a particular emphasis on the linguistic nuances of Odisha and eastern India.

The dataset features recordings from a single native Hindi speaker, ensuring consistency in voice characteristics. The speech samples vary in length, ranging from 9 seconds to over 2 minutes, and encompass a diverse array of content, including formal narration, greetings, traditional expressions, and culturally contextual material.

🗂️ Dataset Structure

id: Unique identifier for each sample (e.g., sample_001)
speech: Filename of the corresponding WAV audio file
text: Transcribed Hindi text content

Dataset will not be available in the dataset viewer as there is no train partition.

🔊 Audio Sampling Rates & Model Compatibility

To cater to a broad spectrum of research and application needs, the dataset provides audio recordings in three different sampling rates:

🎧 16 kHz (Wideband)

Description: Standard sampling rate widely used in speech processing tasks.
Ideal For:
- OpenAI Whisper: Requires 16 kHz input audio for optimal performance.
- Facebook Wav2Vec 2.0: Pretrained on 16 kHz sampled speech audio.
- Mozilla DeepSpeech: Expects 16 kHz mono-channel WAV files.

🎧 24 kHz (High-Quality Wideband)

Description: Offers a balance between audio quality and computational efficiency.
Ideal For:
- Parler-TTS: Supports 24 kHz audio, suitable for high-fidelity TTS applications.
- ESPnet-TTS: Flexible with sampling rates; 24 kHz is commonly used.

🎧 32 kHz (Ultra-Wideband)

Description: Provides higher audio quality, beneficial for advanced TTS models and applications demanding superior clarity.
Ideal For:
- Advanced TTS Models: Models that can leverage higher sampling rates for improved synthesis quality.

Note: When using models that require a specific sampling rate, ensure that the input audio matches the expected rate to avoid potential degradation in performance.

🧪 Applications

Automatic Speech Recognition (ASR): Fine-tuning and evaluating models like Whisper, Wav2Vec 2.0, and DeepSpeech for Hindi language transcription tasks.
Text-to-Speech (TTS) Synthesis: Training and assessing TTS models such as Parler-TTS and ESPnet-TTS for generating natural-sounding Hindi speech.
Speech-Text Alignment: Developing and testing alignment algorithms for synchronizing speech with textual content.
Linguistic Research: Analyzing phonetic and prosodic features of Hindi as spoken in the Odisha region.

🧾 License

This dataset is shared by @BBSRguy as part of the ODEN Initiative for open AI development in India, aiming to democratize AI and make it accessible for everyone, particularly focusing on developing AI tools like personal assistants for the people of Odisha.

🙏 नमस्ते! 🙏