Datasetpublic

eQOURSE/multilingual-speech

Multilingual Indian Conversational Speech A dataset of naturalistic, spontaneous two-speaker conversations across 13 Indian languages, with segment-level transcripts, speaker profiles, timestamps, and recording metadata. Designed for ASR, TTS, speaker diarization, and conversational speech research. Languages (13) Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Nepali, Odia, Punjabi, Tamil, Telugu, Urdu. Content Conversations… See the full description on the dataset page: https://huggingface.co/datasets/eQOURSE/multilingual-speech.

sourceHugging Facecc-by-4.0updated 3mo agoView on Hugging Face

0likes93downloads

Dataset Card

Multilingual Indian Conversational Speech

A dataset of naturalistic, spontaneous two-speaker conversations across 13 Indian languages, with segment-level transcripts, speaker profiles, timestamps, and recording metadata. Designed for ASR, TTS, speaker diarization, and conversational speech research.

Languages (13)

Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Nepali, Odia, Punjabi, Tamil, Telugu, Urdu.

Content

Conversations resemble real-world interactions across multiple domains: Technology / customer support, Financial services, Healthcare, Food & delivery, and Retail & commerce. Speech is conversational and spontaneous — natural turn-taking, code-switching (English mixed with the local language), interruptions, and expressive prosody.

Structure

audio/                    full conversation recordings (WAV), one per language
metadata.jsonl            segment-level annotations referencing the recordings

Each row is one utterance segment (mostly 1–20 s) that references its source recording in audio/ with start/end timestamps.

Schema

Field	Type	Description
`segment_id`	string	Segment identifier within a recording (`SEG-001`).
`recording_id`	string	Source recording ID (`REC-ASM-HLT-011`).
`audio_file`	string	Relative path to the full recording in `audio/`.
`language`	string	Language of the conversation.
`speaker_id`	string	Speaker label (`SPK_01`, `SPK_02`).
`speaker_role`	string	Conversational role (Customer, Agent, Pharmacist, ...).
`speaker_gender`	string	Speaker gender (from speaker profile).
`speaker_age`	string	Age bracket (e.g. `Adult (18+)`).
`speaker_region`	string	Speaker region/location.
`accent_dialect`	string	Accent or dialect description.
`start_time`	string	Segment start (`HH:MM:SS.mmm`).
`end_time`	string	Segment end (`HH:MM:SS.mmm`).
`start_seconds`	float	Segment start in seconds.
`end_seconds`	float	Segment end in seconds.
`duration_seconds`	float	Segment duration in seconds.
`transcript`	string	Verbatim transcript in the native script.
`domain`	string	Conversation domain.
`collection_method`	string	How the audio was collected.
`environment_type`	string	Recording environment description.
`recording_date`	string	Date of recording.
`sample_rate`	int	Audio sample rate (Hz).
`channels`	int	Number of audio channels.
`bit_depth`	int	Audio bit depth.
`source_recording`	string	Original recording filename.

Audio

—Uncompressed WAV, mostly 48 kHz / 16-bit (some 44.1 kHz and 24-bit).
—One full conversation recording per language; segment rows reference offsets within it, so clips can be extracted on demand from start_seconds / end_seconds.

Notes

—Transcripts include natural English code-switching, common in Indian conversational speech.
—Timestamps normalized to a consistent HH:MM:SS.mmm format; a few segments with source timestamp inconsistencies have a null duration.
—Speaker attributes (role, gender, age, region, accent) come from the per-recording speaker profile block in the source annotations.

Extracting a segment clip (example)

python

import soundfile as sf, json
row = json.loads(open("metadata.jsonl").readline())
data, sr = sf.read(row["audio_file"])
clip = data[int(row["start_seconds"]*sr):int(row["end_seconds"]*sr)]
sf.write("segment.wav", clip, sr)