Total number of audio files used to train DeepSpeech 0.5.1

Deep Speech
#1

In the release notes of DeepSpeech 0.5.1, it is mentioned that the model was trained for 467356 steps or 75 epochs. Going by the batch training batch size of 24

steps_per_epoch = 467356 / 75 ≅ 6231
total_number_of_samples = 6231 * 24 = 149544

Did the model achieve a word error rate of 8.22% by training only on 149544 audio samples? And what was the maximum audio length (in seconds) considered while processing the dataset?

#2

wc -l fisher-train.csv librivox-train-{clean,other}-???.csv swb-train.csv

1041522 fisher-train.csv
28540 librivox-train-clean-100.csv
104015 librivox-train-clean-360.csv
148689 librivox-train-other-500.csv
124708 swb-train.csv
1447474 total

Minus one from each file for the CSV header, for a total of 1447469 samples. Maximum duration clipped at 10 seconds.

– reuben

1 Like