Transcription, Forced Alignment and Diarization for Pangloss Collection

This repository provides tools for forced alignment, segmentation, and transcription of Pangloss XML-annotated audio data, using transformers and pyannote pipelines.

Features

• Transcription: Transcribes audio or segments using a pretrained Wav2Vec2 model and diarization.
• Forced Alignment: Aligns words in Pangloss XML files with corresponding audio segments and writes word-level timestamps back to XML.
• Segmentation: Splits audio into speech segments using Speech Activity Detection (SAD).
• XML Parsing & Audio Chunk Extraction: Extracts sentences and corresponding audio chunks from Pangloss XML.

Installation

Linux / macOS

To set up the environment and install the package, follow these steps:

chmod +x setup.sh
./setup.sh

Windows :

1. Accept pyannote/segmentation-3.0, pyannote/speaker-diarization-3.1, and pyannote/voice-activity-detection on Hugging Face.

2. Create an access token at hf.co/settings/tokens.

3. Set your Hugging Face token in the environment variable:

   $env:HF_TOKEN="your_token_here"

   or use the --token argument in the CLI commands.

4. Install dependencies in a virtual environment using pixi:

   pixi install

5. Install the package
   This allows you to use the package as a CLI tool and import it in Python scripts.

pixi run python -m ensurepip --upgrade
pixi run python -m pip install --upgrade pip
pixi run pip install -e .

6. Download or prepare WAV files (and Pangloss XML files) and place them in the data/ directory (recommended).

7. Download or train a Wav2Vec2 model and use it in the --model argument, place ir in models/ directory (recommended).

Usage

All main features are available as CLI commands thanks to the package structure and [project.scripts] entry points.
You do not need to use python ... directly.
Use the following commands from your project root (or with pixi run ...):

Transcribe and segment long audio files and get TextGrid

Uses diarization and ASR to create a TextGrid file handling multiple speaker tiers and distinguishing between human voice and silence/noise.

pixi run transcribe --model models/Na_best_model --audio_path data/235213.wav --num_speakers 1

• Outputs a transcribed TextGrid file.

Forced Alignment and Word-Level Timestamps

pixi run word_align --pangloss_xml data/235213.xml --wav data/235213.wav --model models/Na_best_model

• Outputs an aligned Pangloss XML file with word-level <AUDIO start="..." end="..."/> tags.

Transcribe short audio files and get a .txt file

Simplistic transcription of short audio files using a pretrained Wav2Vec2 model.

pixi run simple_predict data/235213.wav --model models/Na_best_model

• Outputs data/235213.txt containing the transcription.
• Does not handle multiple speakers.

---

For more details, see the docstrings in each module or run:

pixi run transcribe --help
pixi run word_align --help
pixi run simple_predict --help

---

Notes

• All package code is in src/nlp_pangloss/.
• CLI commands are defined in the [project.scripts] section of pyproject.toml.
• For advanced usage or development, you can still run scripts directly with:

  pixi run python src/nlp_pangloss/segment_and_transcribe.py ...

  but this is not necessary for typical use.

---