guranshsaran/soundsense-sudormrf-separation

SoundSense AI - GroupComm SudoRM-RF Speech Separator

This is a fine-tuned/adapted model checkpoint for SoundSense AI.

Base model / source:

• SudoRM-RF (Group Communication variant), https://github.com/etzinis/sudo_rm_rf
• Tzinis et al., "Sudo rm -rf: Efficient networks for universal audio source separation"

Training data:

• LibriSpeech train-clean-360 (speech) + WHAM! noise (5000 real ambient noise recordings)
• Synthetic 2/3-speaker noisy mixtures generated on-the-fly, random SNR -5 to +20 dB

Use:

• Part of SoundSense AI hackathon submission (Stage 2: Speech Separation).
• Isolates up to 3 simultaneous speakers from a noisy mixed-audio input.

Limitations:

• Built for prototype/demo use.
• Current SI-SNR: 2.51 dB (2-speaker noisy), 0.59 dB (3-speaker noisy) — below the target KPI of >18 dB / >10 dB respectively. Separation quality is not yet sufficient for clean speaker isolation; further training required.
• Performance should be verified on the target environment before deployment.