Behnooshn/seamlessm4t-enfr-s2st-eval

SeamlessM4T English-to-French S2ST Evaluation

This dataset contains evaluation outputs from a fine-tuning experiment with Meta SeamlessM4T for English speech → French speech translation.

Experiment

We fine-tuned seamlessM4T_medium on an aligned English-French speech dataset.

Dataset split:

Split	Samples
Train	1,779
Validation	100
Test	100

The audio was converted to 16 kHz mono WAV before training.

Results

Evaluation was performed on the 100-sample held-out test set.

Model	BLEU	chrF	COMET
Base SeamlessM4T	26.24	48.43	0.7751
Fine-tuned SeamlessM4T	## 38.00	## 59.25	## 0.7988

The fine-tuned model improved over the base model on all three metrics.

Files

predictions.tsv              # source, reference, base output, fine-tuned output
metrics_with_comet.txt       # final metric scores
wav_outputs_manifest.tsv     # mapping between test items and generated wavs
base_wavs/                   # S2ST outputs from base SeamlessM4T
finetuned_wavs/              # S2ST outputs from fine-tuned SeamlessM4T