Speech2Text

A simple wrapper around DeepSpeech2 for end‑to‑end speech‑to‑text inference.

Description

This repository demonstrates how to run inference using the DeepSpeech2 model (ASR‑DeepSpeech2) on your own audio files. Follow the steps below to clone, configure, and execute the pre‑trained model.

Prerequisites

• Python 3.7 or higher
• git
• pip
• (Optional) Google Colab or any environment with /content paths (adjust paths if running locally)

How to Run

1. Clone this repo and DeepSpeech2

# Clone your Speech2Text project
git clone https://github.com/yegee0/Speech2Text.git
cd Speech2Text
git checkout lm_beam_search

2. Install dependencies

pip install -r requirements.txt
pip install gdown

3. Download pre‑trained models & resources

# Clean acoustic model
gdown 1NHkGDwOUmSyGqO4zHpbLEOAxokygw6m2

# Alternative acoustic model
gdown 1H_tc1sCFfDFFaElu09DuvXi8aaYnGZif

# Pre‑trained language model
gdown 1-hy2wFlaSU46XhpgEVc0E8lIBcHdOQVq

# Lexicon (pronunciation dictionary)
gdown 1ggUebF7z-k5Fz7mjUWq6NHilV3eHiSb7

4. Prepare your audio & transcripts

# Create folders for your test audio and transcripts
mkdir -p /content/my-audio
mkdir -p /content/my-transcripts

# Copy your .wav file(s), e.g. harvard1.wav
cp /content/harvard1.wav /content/my-audio/

# Create an empty transcript file
echo "" > /content/my-transcripts/harvard1.txt

Note: If you’re running outside of Colab, replace /content/... with your local paths, e.g. ./my-audio and ./my-transcripts.

5. Run inference

python inference.py \
  -cn=inference_clean_local \
  '+datasets.test.audio_dir=/content/my-audio' \
  '+datasets.test.transcription_dir=/content/my-transcripts' \
  '++dataloader.batch_size=1'

• -cn selects the config name (e.g. inference_clean_local).
• Adjust batch_size to control parallelism.
• Transcriptions will be written into your transcripts folder.

Tips & Troubleshooting

• Path Errors: Make sure all directories exist and paths are correct.
• Download Issues: Verify that each gdown command completes; you should have four files in the root of ASR-DeepSpeech2.
• Config Variants: Check configs/ for other -cn options (e.g. noisy or beam‑search variants).
• Batch Size: Increase for parallel processing, but watch GPU/CPU limits.