jssaluja/zipformer-large-causal-sggs-ncer-train-107387-epochs-50-test-1569

This model is a Zipformer transducer (icefall zipformer size large) trained from scratch on private Punjabi Gurbani speech. It achieves the following results on the evaluation set:

• Loss: 2.0341
• Wer: 0.5329
• Wil: 0.7453
• Mer: 0.5183
• Cer: 0.2170
• Ncer: 0.1884
• Ccer: 0.2389

Model description

• Architecture: icefall Zipformer large transducer, 151,203,546 parameters
• train_scope: verse_only
• best_model_metric: ncer (0.18823617788461539 at epoch 48)
• smoke_test: False
• train rows: 107387
• test rows: 1569
• Checkpoint selection uses full-dev greedy decode, not a subsample.

Intended uses & limitations

Punjabi Gurbani and general-Punjabi ASR. Not a transformers CTC checkpoint; load model.pt with the icefall Zipformer used to train it.

Inference

model.pt is the Zipformer state dict and bpe.model is the SentencePiece model trained with it (vocab 1000, blank id 0). This is not a transformers checkpoint, so pipeline() will not load it. Greedy CTC on the auxiliary head is the decode used for the CER numbers above.

pip install torch torchaudio sentencepiece lhotse datasets huggingface_hub soundfile
git clone https://github.com/k2-fsa/icefall

import torch, sentencepiece as spm
from huggingface_hub import hf_hub_download
repo = 'jssaluja/zipformer-large-causal-sggs-ncer-train-107387-epochs-50-test-1569'
weights = hf_hub_download(repo, 'model.pt')
bpe = spm.SentencePieceProcessor(model_file=hf_hub_download(repo, 'bpe.model'))
# build the network with train_zipformer's build_model() from the training
# CLI defaults, then: model.load_state_dict(torch.load(weights, map_location='cpu'))
# 16 kHz mono -> 80-dim Lhotse fbank -> per-utterance mean/std.
# ids = greedy argmax over model.ctc_output(encoder_out), drop blank 0 and repeats.
# text = bpe.decode(ids)

Training and evaluation data

Source collection names are not published.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.045
• max_duration: 600.0
• seed: 20260830
• optimizer: ScaledAdam with clipping_scale=2.0 and Eden (lr_batches=7500, lr_epochs=3.5)
• num_epochs: 50
• mixed_precision_training: Native BF16 AMP

Training results

(table generated directly from epoch-end full-dev history, keyed by metric name -- avoids a transformers model-card column-ordering bug seen in earlier runs. Validation Loss is the training objective on the dev cuts, sum / target tokens, not MMS CTC eval_loss.)

Framework versions

• Pytorch 2.12.0+cu126
• icefall 6fd674312c1d87bd9fc888d623cb3e347ac019ff