SPRING_INX contains rich, diverse, and naturally produced speech, covering a wide range of speakers, acoustic environments, and linguistic styles. A unique hallmark of this corpus is its code-mixed transcriptions, where native-language text naturally incorporates Romanised and English words - a phenomenon common in Indian bilingual communication. The dataset includes both monologue and conversational recordings, creating realistic scenarios suitable for robust, real-world speech system development. Key Features: ~3400 hours of labelled speech (Phase 1: ~2000h, Phase 2: ~1400h) 10 Indian languages with broad linguistic and acoustic coverage (Assamese, Bengali, Gujarati, Malayalam, Hindi, Marathi, Odia, Punjabi, Tamil, Kannada) 16 kHz high-quality audio compatible with ASR & SSL models (e.g., Wav2Vec2, HuBERT, Whisper) Code-mixed transcripts enabling bilingual, transliteration, and multilingual modelling Monologue + conversational speech for realistic system training Large speaker diversity for robust speaker modelling General-purpose domain content with rich semantic variation Official ESPnet recipe for easy and reproducible end-to-end ASR training (https://github.com/espnet/espnet/tree/master/egs2/spring_speech)
© 2026 - Copyright AIKosh. All rights reserved.