This repo is used to identify Chinese Numbers from speech signals.

---

target

Recognition of Chinese eight-digit speech. / 识别中文的八位数字语音。

---

main idea

1. change speech signal to image
2. use CNN recognize image

---

plan A: recognize-single-number (Done)

• input:
  0_a.wav | 0_b.wav | 2_a.wav | 3_a.wav | 3_b.wav | ...
• output:
  0 | 0 | 2 | 3 | 3 | ...
• model：
  VGG13 for 10 class(0-9 numbers)
• data:
  2k+ single number specch data from 60+ people
  use 5 kinds of augment, include crop、loudness、noise、pitch and speed
  final training data 13k+
• pretrained_model

---

plan B: end-2-end (Lack of Data)

• input:
  08416923_a.wav | 79684315_b.wav | 29368741_a.wav | ...
• output:
  08416923 | 79684315 | 29368741 | ...
• model：
  speech recognition has two models: acoustic-model and language-model
  plan-B just use acoustic-model: CNN+LSTM+CTC
• data:
  lack of continuous voice data
• pretrained_model