Deep Speech model with Catalan

DeepSpeech version: 0.7.0-alpha.2

• output_graph.pb

• output_graph.tflite

• output_graph.pbmm

• kenlm.scorer

• best_dev-1298770.meta

• best_dev-1298770.data-00000-of-00001

• alphabet.txt

Training with more Data:

TF_FORCE_GPU_ALLOW_GROWTH=true python -u DeepSpeech.py --noearly_stop --alphabet_config_path alphabet.txt --load_train auto --train_files train.csv --train_batch_size 8 --dev_files dev.csv --dev_batch_size 8 --test_files test.csv --test_batch_size 8 --checkpoint_dir checkpoints/ --train_cudnn --n_hidden 2048 --epochs 40 --learning_rate 0.0001 --export_dir export/ --export_tflite --scorer_path kenlm.scorer --augmentation_pitch_and_tempo_scaling --augmentation_freq_and_time_masking --augmentation_spec_dropout_keeprate 0.8

Predict with audio

(16000 Hz wav format)

deepspeech --model output_graph.pb --scorer kenlm.scorer --audio audio.wav

Latest Training Loss

• Loss: 61.576375
• Avg WER: 0.140000