TO BUILD AND RUN THE APPLICATION USING DOCKER

• Go to root directory
• Run below commands

To Build

docker-compose build

To Run

docker-compose up

---

Bangla Speech-to-Text Pipeline with Meeting Analysis

This architecture outlines a comprehensive solution for processing Bangla audio files and extracting meaningful information from meeting transcriptions.

1. Audio Upload & Validation Layer

• Audio Upload: Users upload audio files (e.g., MP3, WAV, M4A).
• Audio Validation:
  • Validates the uploaded file for size, format, and quality (bitrate, sample rate, channels).
• Audio Quality Check: Ensures the uploaded audio meets the necessary standards for processing.

2. Audio Preprocessing Pipeline

• Audio Segmentation:
  • Splits long recordings into smaller, manageable segments (15-30 seconds).
  • Removes silent periods.
• Noise Reduction:
  • Applies adaptive noise reduction techniques to improve audio quality.
• Speech Enhancement:
  • Uses spectral subtraction to reduce background noise.
• Format Conversion:
  • Converts the audio to a standard 16kHz mono WAV format (if necessary).

3. ASR (Automatic Speech Recognition)

• Primary ASR Model:
  • Option 1: Fine-tune the Whisper model on Bangla data.
  • Option 2: Use a custom ASR model (e.g., Wav2Vec 2.0 or Conformer) trained on Bangla speech.
  • Option 3: Google Speech To Text (Free Trial).
• Speaker Diarization:
  • Segments the audio by speaker and labels the transcribed text accordingly.

4. Post-processing Pipeline

• Text Normalization:
  • Handles special characters, numbers, and ASR errors.
• Language Model Correction:
  • Implements a Bangla-specific language model to handle domain-specific terms and grammatical errors.

5. Analysis Layer

• Text Summarization:
  • Extracts key discussion points and generates concise summaries.
• Action Items Detection:
  • Identifies tasks, deadlines, and responsibilities from the transcript.
• Speaker Attribution:
  • Matches each part of the transcript to the respective speaker.

6. Storage & Database Design

Database

Stores processed data including:

• Meetings: Information about each meeting (e.g., ID, duration, status).
• Transcripts: Full transcriptions, including their processed status.
• Speakers: Speaker labels and corresponding speaking segments.
• Action Items: Tasks identified during the meeting, with assignee and due date.

Example SQL Schema

-- Core Tables

meetings
  - meeting_id
  - date_time
  - duration
  - audio_file_path
  - status

transcripts
  - transcript_id
  - meeting_id
  - full_text
  - processed_status

speakers
  - speaker_id
  - meeting_id
  - speaker_label
  - speaking_segments

action_items
  - item_id
  - transcript_id
  - description
  - assignee
  - due_date

Mermaid Live Editor

Link