Finnish Speech Recognition using LSTM-CTC

A deep learning model for Finnish speech recognition using LSTM networks with CTC (Connectionist Temporal Classification) loss.

Overview

This project implements an LSTM-based speech recognition system specifically trained for Finnish language audio. It utilizes both Kielipankki and Hugging Face datasets for training and evaluation.

Prerequisites

• Python 3.12.7+

Install dependencies

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
pip install -r requirements.txt

exact versions used in training and running locally defined in freeze.txt notice pywin32 commented out in requirements.txt

Directory Structure

data/
├── kielipankki/
│   ├── dev-test/
│   ├── set1-part1/
│   └── set1-part2/
└── hf/
    └── [huggingface_dataset (common_voice_17)]

Kielipankki Data

For all datasets except dev-test, place the inner folder from each part directly into the kielipankki directory.

Hugging Face Data

If not pre-downloaded, you'll need to provide your Hugging Face token during runtime.

Basic Training

python train.py \
    --data /path/to/data/folder \
    --lr 0.001 \
    --epochs 100 \
    --batch_size 32 \
    --num_workers 4

Using Pre-downloaded Data

python train.py \
    --data /path/to/data/folder \
    --predl true \
    --batch_size 32

Using Hugging Face Datasets

python train.py \
    --data /path/to/data/folder \
    --token YOUR_HUGGINGFACE_TOKEN \
    --batch_size 32

Command Line Arguments

Argument	Description	Default
--data	Path to data directory	Required
--token	Hugging Face API token	None
--predl	Use pre-downloaded data	False
--lr	Learning rate	0.001
--epochs	Number of training epochs	100
--batch_size	Batch size for training	32
--num_workers	Number of data loading workers	4

Model Architecture

The model uses a bidirectional LSTM architecture with CTC loss for sequence-to-sequence learning.