You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Dataset

Synthetic two-party phone conversations (agent and customer), one row per spoken turn. Each row holds the turn text, its audio (24 kHz mono, lossless FLAC) and speaker, scenario and generation metadata.

56,606 clips, 144.51 h of audio, 11 push(es).

Language Clips Hours
bn-IN 5,781 14.90
en-IN 5,560 15.24
gu-IN 5,685 14.62
hi-IN 5,495 13.49
kn-IN 5,621 14.45
ml-IN 5,752 14.52
mr-IN 5,761 14.81
pa-IN 5,654 13.79
ta-IN 5,628 14.23
te-IN 5,669 14.47
Gender Clips Hours
female 39,361 100.67
male 17,245 43.84

Loading

from datasets import load_dataset

ds = load_dataset("saurabh1043/asset-probe-00cd7fbb", split="train")                    # all languages
hi = load_dataset("saurabh1043/asset-probe-00cd7fbb", "hi-IN", split="train", streaming=True)
speakers = load_dataset("saurabh1043/asset-probe-00cd7fbb", "speakers", split="train")
conversations = load_dataset("saurabh1043/asset-probe-00cd7fbb", "conversations", split="train")

Without datasets: every file under data/ is plain Parquet; audio.bytes is a complete FLAC file (soundfile.read(io.BytesIO(row['audio']['bytes']))).

Layout

Path Content
data/<language_code>/<chunk_id>-<batch_id>-<nnn>.parquet Clips, sorted by conversation and turn.
metadata/speakers/<batch_id>.parquet Designed voices: prompt, gender, age, city, preview audio (one file per batch or campaign).
metadata/conversations/<batch_id>.parquet Scenario, business facts, full message list, outcome.
metadata/index/<chunk_id>.parquet Utterance id, source WAV sha256 and file of every clip in a push.
metadata/chunks.jsonl One line per push: clips, hours, breakdowns, files.

Clip columns

Column Meaning
audio {bytes, path}: FLAC, 24 kHz mono 16-bit.
utterance_id, conversation_id, scenario_id, turn_index Ids; turn_index orders a conversation.
batch_id, chunk_id Generation run and push the clip came in.
role assistant (agent) or user (customer).
language_code, language_name, text, word_count Language and the exact text read.
speaker_id, gender, speaker_age, speaker_city, speaker_cross_lingual Speaker; ids are stable across languages and batches.
duration_sec, speech_duration_sec, sample_rate File length and length without edge silence.
domain, workflow, code_mix_level, script_style, register, customer_mood, emotion, length_bucket Scenario and style labels.
named_entities [{type, value}] spoken in the turn.
quality_status, utterance_warnings, audio_warnings Text validation status and audio checks (DURATION_OUTLIER, NEAR_SILENT).
tts_model, tts_service_tier, tts_style, voice_id, voice_language_code, voice_prompt_sha256 How the clip was synthesized.
audio_sha256, generated_at sha256 of the source WAV and synthesis time.

Pushes

Chunk Batch Clips Hours Pushed
c0001 xlang01-heldout 3,129 5.67 2026-10-27T06:12:58+00:00
c0002 xlang01-a2 4,609 12.02 2026-10-27T06:16:07+00:00
c0003 xlang01-a1 5,683 15.02 2026-10-27T06:19:36+00:00
c0004 xlang01-a1 5,501 15.02 2026-10-27T06:22:35+00:00
c0005 xlang01-a1 5,809 15.01 2026-10-27T06:25:25+00:00
c0006 xlang01-a1 5,858 15.00 2026-10-27T06:28:01+00:00
c0007 xlang01-a1 5,835 15.02 2026-10-27T06:30:48+00:00
c0008 xlang01-a1 5,735 15.01 2026-10-27T06:33:05+00:00
c0009 xlang01-a1 6,022 15.01 2026-10-27T06:35:48+00:00
c0010 xlang01-a1 5,842 15.02 2026-10-27T06:38:14+00:00
c0011 xlang01-a1 2,583 6.71 2026-10-27T06:40:19+00:00
Downloads last month
5,811