README.md

Podcast highlight extraction

Installation

The repo was developed on MacBook M1 chip laptop. Thus, Python 3.8 and most packages, as well as the virtual environment, are managed using Conda. The list of packages is shown in the file requirements.txt.

Problem

sentencepiece might not be able to be import on Apple M1

arch -arm64 brew install cmake pip install --no-cache-dir sentencepiece

Data

Sound data are stored under the directory ./sound

Step 1: Speaker segmentation (find_speaker_segment.py)

Using SpeechBrain's speaker embeddings from the HuggingFace repo of "spkrec-ecapa-voxceleb". See https://huggingface.co/speechbrain/spkrec-ecapa-voxceleb

Step 2: Audio tagging by AudioSet (assign_audioset_labels.py)

• The sound oncology is shown in the following link: https://research.google.com/audioset/ontology/index.html
• Vggish and YAMNet Download those two models for AudioSet from a directory within TensorFlow repo (Trick: DownGit) https://github.com/tensorflow/models/tree/master/research/audioset

Step 3: Find highlight through the combination of sentence embedding (BERT embeddings) and music presence

• Huggingface's sentence transformer https://huggingface.co/sentence-transformers/sentence-t5-large
• Final score = Alpha * Music Score (0 or 1) + (1 - Alpha) * Sentence similarity