Fine-tuning corpus for full-duplex multi-speaker dialogue.
data_{zh,en}{,_multichannel}.jsonl (one record per line):
field
type
description
path
string
relative path to the audio file
voice
string
relative path to speaker prompt
duration
float
clip duration in seconds
system
string
persona / system prompt
transcripts/*.parquet:
column
type
description
audio_path
string
matches data_*.jsonl path
idstring
duration… See the full description on the dataset page:
https://huggingface.co/datasets/MultiTalk/MultiTalkFT.
