NOTSOFAR-1: Single-Device Far-Field Meeting Transcription Dataset is a dataset released by Microsoft in 2025 for evaluating far-field meeting transcription tasks, aiming to advance the development of single-device far-field speech separation and transcription technologies.

This dataset contains real meeting recordings, covering various meeting rooms, numbers of speakers, microphone configurations, and acoustic environments. It can be used for research on far-field speech acquisition, multi-speaker overlapping speech processing, and meeting transcription. The official release also includes baseline system code, integrating Continuous Speech Separation (CSS), Whisper-based Automatic Speech Recognition (ASR), and NeMo-based speaker diarization modules, enabling researchers to run inference pipelines and evaluate model performance.

Dataset Composition

• Development Set (Dev-set): Contains single-channel and multi-channel meeting recording data. There are 177 single-channel sessions and 106 multi-channel sessions.
• Subset Splits: Based on meeting scenarios and noise conditions, the data is divided into multiple subsets, including Natural Meetings (#NaturalMeeting), Debate Overlaps (#DebateOverlaps), Non-overlapping Turns (#TurnsNoOverlap), High Transient Noise (#TransientNoise=high), and Speaking Near Whiteboard (#TalkNearWhiteboard), among others.
• Baseline System Code: Provides complete Python baseline code, including continuous speech separation, Whisper-based ASR, and NeMo-based speaker diarization modules.
• Evaluation Metrics: Primarily uses speaker-attributed tcpWER for ranking, with tcORC-WER serving as an auxiliary analysis metric.
• Training Data: In addition to the development set, a Simulated Training Set is provided, containing approximately 1,000 hours or 200 hours of simulated data for model training and fine-tuning.