Numo Indic Speech
Scripted, single-speaker read speech across Bengali, Hindi, Tamil, and Telugu: 353,349 recordings totaling 3,845.5 hours, published as one configuration per language. Every recording carries its reference transcript, speaker metadata, and automated validation metrics, with full-resolution WAV audio embedded in typed Parquet shards.
IndicSingle-speakerTranscripts
Overview
- 1.1
More than 3,800 hours of scripted read speech across four Indic languages
- 1.2
Transcripts keep the code-switching and English terms present in the original prompts
- 1.3
Automated error-rate and synthetic-speech screening on every recording
Specifications
- License
- CC BY 4.0
- File types
- parquet
- Audio codec
- PCM
- Sampling rate Hz
- 48,000
- Bit depth
- 16
- Channels
- 1 / 2
- Bitrate kbps
- 768 / 1536
- Transcript source
- Ground Truth
- Speaker format
- Single-speaker recordings
Metadata coverage
audio identifierslanguage codescollection topicdurationsample ratechannelsaudio formattranscript textspeaker identifiersspeaker agespeaker genderspeaker countryspeaker regionword error ratecharacter error ratesynthetic-speech suspicion scoreSHA-256 checksums
