2# AI Great Challenge Preliminary Competition Round 1

本文件是為AI語音大擂台的準備文件，為了描述整體程式的架構跟所用到每個Component。

Figure 1: Flowchart of 第一版程式

Figure 1 是整個main.py的流程，

  1. 導入要用到python modules
  2. 設定好 Google Speech-to-Text Service
    • Client 是一個本地端程序(procedure?)，負責把程式的指令傳送到Google Server[?]
  3. 使用glob.glob()取得指定資料夾(./BC/C/)裡面所有副檔名為.wav的檔案名稱
  4. 使用for loop，對於每個檔案，把對應的audio匯入
  5. 一個audio包含4個選項，丟入Sentence Segmentation後把4個選項的音頻分開
  6. 每個選項就是一個句子，呼叫Google Speech API把audio轉換成Text
  7. 對於選項1~4，計算每個選項的字數，我們選擇字數最少的選項為答案，如果有兩個以上有最少的字數，則選擇選項數字最小的那一個
  8. 把每個問題的答案存成out.csv，上傳到Kaggle。
1   #!/usr/bin/env python3
2   # -*- coding: utf-8 -*-
3   """
4   Created on Sat Jun 23 08:57:19 2018
5   
6   @author: PeterTsai
7   """
8   
9   from pydub import AudioSegment
10  from pydub.silence import split_on_silence
11  
12  # [START speech_quickstart]
13  import io
14  import os
15  import numpy as np
16  import pandas as pd
17  import glob
18  
19  # Imports the Google Cloud client library
20  # [START migration_import]
21  from google.cloud import speech
22  from google.cloud.speech import enums
23  from google.cloud.speech import types
24  # [END migration_import]
25  
26  # Instantiates a client
27  # [START migration_client]
28  client = speech.SpeechClient()
29  # [END migration_client]
30  
31  config = types.RecognitionConfig(
32      encoding=enums.RecognitionConfig.AudioEncoding.LINEAR16,
33      sample_rate_hertz=16000,
34      language_code='cmn-Hant-TW')
35      
36  speechFileList = sorted(glob.glob('./BC/C/*.wav'))
37  answer_per_Q = np.zeros(len(speechFileList))+3
38  
39  q = 1 
40  ID = []
41  for speech in speechFileList:
42      print('Question {}'.format(q))
43      song = AudioSegment.from_wav(speech)
44  
45      #split track where silence is 1 seconds or more and get chunks
46  
47      chunks = split_on_silence(song, 
48          # must be silent for at least 1 seconds or 1000 ms
49          min_silence_len=1000,
50  
51          # consider it silent if quieter than -16 dBFS
52          #Adjust this per requirement
53          silence_thresh=song.dBFS
54      )
55  
56      word_num_per_choice = np.zeros(len(chunks))
57      #Process each chunk per requirements
58      for i, chunk in enumerate(chunks):
59          #Create 0.5 seconds silence chunk
60          silence_chunk = AudioSegment.silent(duration=500)
61  
62          #Add  0.5 sec silence to beginning and end of audio chunk
63          audio_chunk = silence_chunk + chunk + silence_chunk
64  
65          # Loads the audio into memory
66          audio = types.RecognitionAudio(content=audio_chunk.raw_data)
67  
68          # Detects speech in the audio file
69          response = client.recognize(config, audio)
70  
71          for result in response.results:
72              print('Choice {}, Transcript: {}, word number: {}'.format(i+1, result.alternatives[0].transcript, len(result.alternatives[0].transcript)))
73              word_num_per_choice[i] = len(result.alternatives[0].transcript)
74  
75      print('word_num_per_choice: ', word_num_per_choice)
76      answer = np.argmin(word_num_per_choice)+1
77      if answer >= 1 and answer <= 4:
78          answer_per_Q[q] = answer
79  
80      ID.append(q)
81      q = q + 1
82  
83  answers = {'ID':ID, 'Answer':answer_per_Q}
84  print(answers)
85  df = pd.DataFrame(data=answers, dtype=np.int32)
86  print(df)
87  df.to_csv('out.csv',index=False)

results matching ""

    No results matching ""