arda-argmax/simchoir

FastMSS synthetic multi-speaker meetings

Synthetic multi-speaker conversational audio generated with FastMSS. Each split contains mixture WAVs (16 kHz, mono), Lhotse manifests (recordings / supervisions / cuts), and per-mixture RTTM files with word-level speaker labels.

Splits / subfolders

• debug/ — 1 mixtures, 1.6 min total, 6 unique speakers
• v0.1/ — 1000 mixtures, 1546.0 min total, 40 unique speakers

Per-split layout

<split>/
    audio/<recording_id>.wav           # 16 kHz mono mixture
    manifests/
        synth-*-train-recordings.jsonl.gz
        synth-*-train-supervisions.jsonl.gz
        synth-*-train-cuts.jsonl.gz   # the file you want for training
        all_cuts.jsonl.gz             # source utterances used by the sim
        all_rooms.json                # RIR room metadata
        noise_files.txt               # noise files used
    rttm_word/<recording_id>.rttm     # word-level SPEAKER lines
    sim.log                           # generator log

Loading

from huggingface_hub import snapshot_download
from lhotse import CutSet
import os

local = snapshot_download("<user-or-org>/<repo-name>", repo_type="dataset")
os.chdir(os.path.join(local, 'v0.1'))  # or 'debug'
cuts = CutSet.from_file('manifests/synth-librispeech-train-cuts.jsonl.gz')
for cut in cuts:
    audio = cut.load_audio()  # path is resolved relative to cwd
    ...

Generation

See the FastMSS repo for the exact simulator config. Each split contains its own sim.log for the full generator output.