liepa3.zipformer-ctc.v01

Zipformer-CTC trained on Liepa3 Corpus

Information in Lithuanian could be found here.

The language resource for the development of AI solutions.

Corpus

Liepa3. 450h

Model results

Demo Jupyter scripts

For short files up to 1m

Script: notebooks/transcribe_zipformer.ipynb <a href="https://colab.research.google.com/#fileId=https%3A//huggingface.co/vdu-liepa3/liepa3.zipformer-ctc.v01/blob/main/notebooks/transcribe_zipformer.ipynb" target="_parent"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/></a>

For long files (with VAD)

Script: notebooks/transcribe_zipformer_vad.ipynb <a href="https://colab.research.google.com/#fileId=https%3A//huggingface.co/vdu-liepa3/liepa3.zipformer-ctc.v01/blob/main/notebooks/transcribe_zipformer_vad.ipynb" target="_parent"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/></a>

Decoder type

Type: Zipformer-CTC.

Training scripts: https://github.com/airenas/icefall/tree/master/egs/liepa3/ASR.

Training params:

./zipformer/train.py --world-size 1 --num-epochs 30 --start-epoch 0 --use-fp16 1 \
	--exp-dir data/exp/v02 \
	--use-cr-ctc 1 --use-ctc 1 --use-transducer 1 --use-attention-decoder 0 --num-encoder-layers 2,2,4,5,4,2 --feedforward-dim 512,768,1536,2048,1536,768 --encoder-dim 192,256,512,768,512,256 --encoder-unmasked-dim 192,192,256,320,256,192 --ctc-loss-scale 0.1 --enable-spec-aug 0 --cr-loss-scale 0.02 --max-duration 400

Export params:

./zipformer/export-onnx-ctc.py \
	--tokens data/exp02/lang_bpe_500/tokens.txt \
	--exp-dir data/exp02/exp/v02 \
	--use-averaged-model 1 \
	--epoch 30 \
	--avg 10 \
	--query-head-dim 32 \
	--value-head-dim 12 \
	--decoder-dim 512 \
	--joiner-dim 512 \
	--causal False \
	--chunk-size 16 \
	--use-cr-ctc 1 --use-ctc 1 --use-transducer 1 --use-attention-decoder 0 --num-encoder-layers 2,2,4,5,4,2 --feedforward-dim 512,768,1536,2048,1536,768 --encoder-dim 192,256,512,768,512,256 --encoder-unmasked-dim 192,192,256,320,256,192 \
	--left-context-frames 128

---