FastMSS: Fast Multi-Speaker Simulation
A toolkit that generates realistic multi-speaker meeting audio from room-impulse-response simulation and HMM-based turn-taking models, producing per-speaker reference signals for training speech separation, enhancement and diarization systems.
FastMSS simulates multi-speaker meetings by combining room impulse response simulation with a hidden Markov model of conversational turn-taking, modeling four transition types: turn hold, turn switch, interruption and backchannel. Transition probabilities can be fit to real meeting corpora such as AMI, CallHome and NOTSOFAR-1, or left at flat priors, and an overlap factor can be tuned to raise or lower interruption and backchannel rates.
Room simulation uses pyroomacoustics with configurable geometry and reverberation time, and noise can be mixed in from corpora such as WHAM or MUSAN at set signal-to-noise ranges. FastMSS reads source speech from Lhotse CutSets with word-level alignments and writes Lhotse-compatible manifests, along with optional RTTM files and NeMo-style diarization manifests.
It can save individual anechoic and reverberant speaker streams alongside the mixed meeting, for training separation and enhancement models as well as speaker-attributed ASR and diarization systems. It scales close to linearly with the number of parallel workers on a DGX node.