Auto Podcast Timestamper

A service that gives timestamped highlights of an audio clip

Uses:

• Mozilla DeepSpeech for VAD and speech-to-text translation
• Microsoft ProphetNet for text summarization

Installation

• sudo apt install sox ffmpeg

Install whatever version of PyTorch suits your system

• pip install torch==1.5.0+cpu -f https://download.pytorch.org/whl/torch_stable.html
• pip install -r requirements.txt

Download models

• mkdir models/
• curl -L https://github.com/mozilla/DeepSpeech/releases/download/v0.7.3/deepspeech-0.7.3-models.pbmm --output models/deepspeech-0.7.3-models.pbmm
• curl -L https://github.com/mozilla/DeepSpeech/releases/download/v0.7.3/deepspeech-0.7.3-models.scorer --output models/deepspeech-0.7.3-models.scorer
• python download_models.py

Usage

Run as Flask API

• export FLASK_APP=main.py
• flask run

API

Request

POST /request

Parameters:

• url: A download URL from Acast for the podcast
• minute_increment: How often to timestamp

curl -X POST -d '{"url": "<URL>", "minute_increment": "<INT>"}' -H 'Content-Type: application/json' http://localhost:5000/request

Response

"{request_id: <UUID>}"

Stream the response back

GET /stream/<request-id>

curl -X GET http://localhost:5000/stream/<request-id>

Response

Either let it stream to your terminal or handle with JS Stream API