DIT Diarization and Transcription tool
DIT.DaT is an experimental, modular pipeline for speaker diarization and automatic transcription of long-form audio, designed to support students and researchers working on spoken-language data.
The project was developed to process very long recordings (e.g. multi-hour interviews and podcasts), preserving speaker separation, producing revision-friendly outputs, and minimizing—but not eliminating—the need for manual intervention.
DIT.DaT was developed by Gabriele Carioli for the Department of Interpretation and Translation (DIT) of the University of Bologna.
The original use case involved academic podcasts and interviews lasting three hours or more, where commercial cloud-based ASR services were either too costly, opaque, or unsuitable for long-term academic workflows. The project therefore focuses exclusively on open-source AI models that can be executed on local hardware.
- Speaker diarization for long audio files (hours-long)
- Robust handling of diarization drift via block-based processing
- Semi-automatic and experimental automatic speaker mapping across blocks
- Two-stage segment cleanup to reduce fragmentation
- High-quality multilingual transcription using Whisper
- Outputs optimized for manual revision using subtitle editors
- Fully persistent, restartable, and inspectable processing pipeline
DIT.DaT combines two mature open-source systems with complementary strengths:
- Used for speaker diarization ("who speaks when")
- Model:
speaker-diarization-community-1 - Supports exclusive mode to handle overlapping speech
- Produces speaker embeddings reused for cross-block speaker mapping
- Used for automatic speech recognition (ASR)
- Model:
large-v3 - Multilingual, end-to-end encoder–decoder architecture
- Particularly effective on long-form and academic speech
- GPU strongly recommended
- The pipeline is designed to run on dedicated workstations
- CPU-only execution is technically possible but impractical for long audio
No specific GPU brand is enforced, but adequate VRAM is required for large models.
- Modular pipeline, not a monolithic application
- Explicit handling of known model limitations
- Human-in-the-loop where necessary and meaningful
- Full transparency via intermediate artifacts
- Reproducibility through versioned configuration
The system is intentionally designed so that every phase can be inspected, corrected, or re-run independently.
Processing is organized into a sequence of independent phases, all operating inside a dedicated project directory.
Each phase reads the output of the previous one and produces persistent artifacts.
- Audio conversion to WAV (16 kHz, mono)
- Optional splitting into fixed-duration blocks (default: 30 minutes)
- Independent diarization of each block
- Extraction of speaker audio samples for manual identification
- Applies the speaker mapping across blocks
- Preserves original segment identifiers for traceability
- Merges consecutive segments from the same speaker
- Removes very short residual segments
- Allows manual removal or merging of speakers
- Generates one audio track per speaker
- Useful for analysis or downstream processing
- Speaker-specific segments passed to Whisper
- Padding added to improve recognition on short segments
- Produces both raw and aggregated transcripts
- Generates SRT subtitles and text files
- Outputs compatible with SubtitleEdit and similar tools
When block-based processing is enabled, speakers are identified independently in each block.
DIT.DaT therefore provides:
- Manual speaker mapping via editable text files
- Experimental automatic mapping based on speaker embeddings
Automatic mapping can significantly reduce manual work but still requires verification.
Final outputs include:
- JSON files with full transcripts and metadata
- SRT subtitle files for revision
- Text files with and without timestamps
- Clean text files suitable for publication
- Optional HTML reports for diarization analysis
Limitations are largely intrinsic to the underlying models:
- Output quality depends strongly on audio quality
- Reduced accuracy with strong regional accents
- Difficulty distinguishing very similar voices
- Imperfect handling of music and sound effects
As a result, manual revision is always expected.
DIT.DaT aims to minimize, not eliminate, human intervention.
This project is released under the MIT License.
Experimental / research prototype
The tool is actively usable but not designed as a turnkey consumer application. It is intended for technically competent users in academic or research contexts.