Vartalaap is a corpus of synthetic Indian customer-support calls whose speaker timbres are cloned from consented research recordings. Access is reviewed individually.

By requesting access you agree to use this dataset for non-commercial research only, to make no attempt to identify or re-synthesise the speakers whose voices informed it, and not to redistribute it.

Vartalaap — full-duplex Hindi/English conversational speech

Dual-channel synthetic Indian customer-support calls for training full-duplex speech-to-speech models.

179,808 calls · 5,411.0 hours · 4942 shards (last updated 2026-11-03 12:57 IST)

Audio layout

Each row's audio is a stereo FLAC at 24000 Hz:

channel	content
0 (LEFT)	agent — pristine, TTS speech and silence only
1 (RIGHT)	user — the caller

from datasets import load_dataset
ds = load_dataset("kapturecx/Vartalaap", split="train", streaming=True)
row = next(iter(ds))
agent = row["audio"]["array"][:, 0]     # left
user   = row["audio"]["array"][:, 1]    # right

Speech generation and engine provenance

From the 2026-10-30 generation configuration, new calls mix drama-3 and FlowMaya. FlowMaya uses the Maya backend through a FlowTTS-compatible websocket API. Its canonical dataset identifier is flowmaya; it is distinct from the historical flowtts engine. Both engines can voice either channel.

`engine_combination`	`agent_engine` (left)	`caller_engine` (right)	New-call sampling target
`d3_d3`	`drama3`	`drama3`	25%
`d3_flowmaya`	`drama3`	`flowmaya`	25%
`flowmaya_d3`	`flowmaya`	`drama3`	25%
`flowmaya_flowmaya`	`flowmaya`	`flowmaya`	25%

These are sampling targets, not measured fractions of the complete dataset. Availability, retries and QA affect the accepted mix. The three engine columns describe each row; voice IDs are speaker identities, not engine identifiers. New FlowMaya audio arrives as 24 kHz PCM16; the stored call remains 24 kHz stereo FLAC with the existing channel layout. Hindi/Hinglish and Indian-English calls are supported. Word timing comes from the corpus alignment stage, not this TTS API.

FlowMaya agents use eight persistent project clones (four male and four female), with stable IDs starting s2s-agent-. They are enrolled once and reused; routine call cleanup never deletes them. Primary callers use temporary usr-s2s- clones created per call and deleted afterward, including failed calls. Their bank has 63 separate reference speakers, excluding the eight agent references. References are complete 3–10 second mono clips with the corresponding exact transcripts; clips are not cropped while retaining a longer transcript. Cleanup after a crash only targets this host's caller clones whose owner process has exited.

Older rows retain their original engine labels and synthesis conditions. This addition does not relabel or regenerate historical audio. In particular, shards 2299–2398 used drama-3 behind an OmniVoice-compatible facade but recorded omnivoice; that historical naming limitation remains. Use drama3 and flowmaya on newer rows and keep this caveat in engine-specific analyses.

Conversation topics and lengths

The 2026-10-30 generation configuration adds general, open-topic conversations alongside customer support. The requested-call mix is 60% general conversation and 40% support. General topics include everyday life, cooking, travel, hobbies, learning, science, stories and friendly debates. A custom topic or topic pool can be supplied; otherwise the generator combines a topic catalog with freshly chosen topics. Generic conversations do not require an account, order or reference ID.

Requested duration range	New-call sampling target
45–120 seconds	50%
3–5 minutes	30%
8–10 minutes	20%

These are configuration targets per requested call, not measured shares of the published corpus or audio hours. Accepted calls must meet their selected duration range and the existing audio/alignment QA gates. New calls are capped at 600 seconds of actual audio, including the mixer tail. A directly requested duration uses a 90–100% acceptance interval to allow a natural ending. Historical rows keep their original durations and generation conditions.

Long conversations are generated in contextual continuations with one topic, speaker cast and caller clone throughout. Duration is checked on rendered speech and the duplex timeline; calls are not padded to length. A short closing is generated separately, with whole trailing body turns removed when necessary to fit it. Barge-ins, listener backchannels and timed overlap remain active across the conversation; random early hangups and service holds are disabled for this duration-controlled mode. This remains synthetic duplex scheduling, with the existing limitations of separately synthesized utterances.

Use domain == "general_conversation" to select generic conversations, scenario for their topic/context, and duration_s for measured audio length. The agent/user role names still identify the left/right channels even when the participants are friends, peers or an assistant and a curious user. Engine fields retain drama3 and flowmaya exactly as documented above. Detailed requested length, segment and topic provenance is retained in per-call JSON and the local manifest; it is not a new Parquet column or a label retrofitted to older rows.

Text conventions

Hindi is written in Devanagari, English in Latin, in the same string — "आपका order अभी out for delivery है". Romanized Hindi is rejected by a QA gate. Everything the TTS spoke is spelled out: no digits (₹4,500 → चार हज़ार पाँच सौ रुपये), abbreviations letter-by-letter in Devanagari (OTP → ओ टी पी), phone numbers in two-digit batches (अट्ठानवे चालीस पैंतालीस बावन तिरासी). Backchannels are deliberately Latin (ji, hmm ji, haan ji).

Fields

call_id, audio, duration_s, sample_rate, channel_layout
domain, scenario, outcome, register, language
agent_voice / agent_gender / agent_speaker_key, same for user_*, plus the full speakers list with per-speaker rate, pitch and target F0
transcript — list of turns: role, kind, start, end, text, emotion, truncated
words_agent / words_user — word-level timings (absolute seconds)
events — duplex annotations: barge_in, agent_yield, user_yield, backchannel, overlap, latency, silence, hold, hangup
n_interruptions, n_backchannels, overlap_ratio, silence_ratio
augmented / augmentations_json — acoustic degradation applied, if any

What makes it full-duplex

Barge-ins land at word onsets and the interrupted side stops within ~60-200 ms, with transcript[i].truncated and the shortened text marking exactly what had been said when the floor was surrendered. Backchannels overlap the speaker without taking the floor. Response latency is drawn per role.

Caveats

Synthetic speech carries TTS artefacts; mix with real audio. Speaker identity is a small set of cloned voices with pitch/rate variation, so treat speakers as acoustic conditions rather than distinct people. augmented=false rows are clean by design — the acoustic degradation pass is separate.

Downloads last month: 1,007

Total file size:

462 GB