IOTEverythin/voxi-duo-synth

Voxi-Duo synthetic recruiter calls

The training and evaluation sets behind IOTEverythin/voxi-duo-en-in-v2: two-channel recruiter calls in Indian English, written from templates and spoken by Chatterbox TTS. Agent on the left channel, caller on the right (telephone band), 24 kHz.

Set	Scripts	Agent voice	Train calls kept	Hours	Eval calls
v0.1	formal	GLOBE, screened through the codec	199	3.21	20
v0.2	spoken English	GLOBE caller_05	197	3.30	20
v0.3	spoken English	Svarah female_01	197	2.97	20

Layout

Each vX.Y/ folder has train.tar and eval.tar (FLAC audio, Whisper transcripts with timestamps as .json, and truth/ with the scripted facts and timing events), plus dialogues.jsonl (the scripts) and the QC report. voices/ holds the reference clips the voices were copied from.

Unpack with ship_dataset.py unpack from the project, or plain tar xf.

How it was made

1. Templated scripts for screening, scheduling and rescheduling calls, with timing events (backchannels, interruptions, corrections, pauses) and a tone per agent line. v0.1 scripts are formal; v0.2 and v0.3 are spoken English.
2. Chatterbox TTS (MIT), voice copied from a reference clip. Its output carries an inaudible watermark.
3. Transcripts from Whisper large-v3; calls whose transcript strayed from the script (WER above threshold, or looped phrases) were dropped.

Licenses and credits

• Caller reference clips: GLOBE (CC0, from Mozilla Common Voice), speakers with a self-reported Indian accent.
• Agent reference clip for v0.3: Svarah by AI4Bharat (CC-BY-4.0).
• The scripts, audio and annotations here: CC-BY-4.0.