A searchable audio transcript interface using Wav2Vec2, Elasticsearch, Flask, and React.

URL: http://3.90.182.103/

Project Structure

• asr/: ASR microservice with wav2vec2 model
• deployment-design/: Architecture design (PDF)
• elastic-backend/: Elasticsearch indexing setup
• search-ui/: Frontend search interface

Run ASR API with Docker

1. Navigate to directory

cd asr

2. Build the image:

docker build -t asr-api ./asr

3. Run the container:

docker run -p 8001:8001 asr-api

4. Test the API:

curl -F "file=@/path/to/sample.mp3" http://localhost:8001/asr

Run Elasticsearch Backend

1. Navigate to directory

cd elastic-backend

2. (Optional but recommended) Create a virtual environment:

python3 -m venv venv
source venv/bin/activate

3. Install Python dependencies:

pip install -r requirements.txt

4. Start Elasticsearch cluster

docker compose up

Open http://localhost:9200/_cat/nodes?v in your browser — you should see both es01 and es02 nodes listed.

5. Index data

python cv-index.py

6. Start backend API server

python search_api.py

Run Frontend (search-ui)

1. Navigate to the frontend directory:

cd search-ui

2. Install dependencies:

npm install

3. Start development sever

npm start

Open http://localhost:3000 with your browser to see the result.

Limitations

• The ASR model used (wav2vec2-large-960h) may produce inaccurate transcriptions, especially for non-US accents or noisy audio, therefore the search function searches both transcribed and actual audio
• Some metadata fields (e.g., age, gender, accent) may be missing or inconsistent in the source CSV file.
• The search UI currently does not support fuzzy matching or partial phrase queries.
• Facets are limited to a fixed number of values (e.g., only top 10 accent types are shown).
• Backend and search functionality assumes the local Elasticsearch and ASR services are running on ports 9200 and 8001 respectively.