Language Identification

This is a template of the lid1 recipe for ESPnet2. It follows a classification-based training/inference pipeline for spoken language identification. The model is trained as a closed-set classifier over a predefined set of language labels. Optionally, language embeddings can be extracted and used for downstream analysis, e.g., t-SNE visualization.

Table of Contents

• Language Identification
  • Table of Contents
  • Recipe flow
    • 1. Data preparation
    • 2. Speed perturbation (Optional)
    • 3. Wav format
    • 4. Statistics collection
    • 5. LID training
    • 6. Inference and embedding extraction
    • 7. Score calculation
    • 8. t-SNE visualization
    • 9-10. (Optional) Pack and upload results
  • How to run
    • Example: VoxLingua107 training

Recipe flow

lid1 recipe consists of 10 stages.

1. Data preparation

Prepares Kaldi-style data directories using local/data.sh.

Expected files include:

• wav.scp: path to raw audio
• utt2lang: utterance-to-language mapping
• lang2utt: language-to-utterance mapping (for sampling)
• segments (optional): used to extract segments from long recordings

1. Speed perturbation (Optional)

Applies offline speed perturbation to the training set using multiple speed factors, e.g., 0.9 1.0 1.1.

1. Wav format

Formats the audio to a consistent format (wav, flac, or Kaldi-ark) and copies necessary metadata to the working directory. Required for both training and evaluation sets.

1. Statistics collection

Collects input feature shape statistics and language information needed for batching and model configuration.

1. LID training

Trains the language identification model using the configuration provided via --lid_config and optional arguments in --lid_args. The model is trained to predict the correct language ID for each utterance.

1. Inference and embedding extraction

Performs inference on evaluation sets. This stage supports both:

• LID prediction (predicted utt2lang)
• Language embedding extraction (utterance-level or averaged per language)
• Optionally saves intermediate outputs

1. Score calculation

Computes standard classification metrics (Accuracy, Macro Accuracy) by comparing model predictions with reference utt2lang.

1. t-SNE visualization

Visualizes the per-language embeddings using t-SNE.

9-10. (Optional) Pack and upload results

Packs the trained model and metadata into a zip file and optionally uploads it to Hugging Face Hub for sharing and reproducibility.

How to run

Example: VoxLingua107 training

Move to the recipe directory:

cd egs2/voxlingua107/lid1

Edit the following files:

vim db.sh        # set path to VoxLingua107 dataset
vim cmd.sh       # job scheduling command if using a cluster
vim conf/mms_ecapa_bs3min_baseline.yaml  # model and training configuration (default training configuration)

Then run the full pipeline:

./run.sh

This will go through all the stages from data preparation to scoring.