mirror of
https://git.roussel.pro/telecom-paris/pact.git
synced 2026-02-09 02:20:17 +01:00
168 lines
5.8 KiB
Python
168 lines
5.8 KiB
Python
import librosa
|
|
import os
|
|
from matplotlib.patches import ConnectionPatch
|
|
import matplotlib.pyplot as plt
|
|
import numpy as np
|
|
import scipy.spatial.distance as dist
|
|
import pyaudio
|
|
import wave
|
|
def dp(distmat):
|
|
N,M = distmat.shape
|
|
# Initialisons the cost matrix
|
|
costmat =np.zeros((N+1,M+1))
|
|
for i in range (1,N+1):
|
|
costmat[i,0]=np.inf
|
|
for i in range (1,M+1):
|
|
costmat[0,i]=np.inf
|
|
|
|
for i in range (N):
|
|
for j in range (M):
|
|
#on calcule le cout minimal pour chaque chemin.pour atteindre the costmat[i][j] il y a trois chemins possibles on choisit celui de cout minimal
|
|
penalty = [
|
|
costmat[i,j], # cas T==0
|
|
costmat[i,j+1] , # cas T==1
|
|
costmat[i+1,j]] # cas T==2
|
|
ipenalty = np.argmin(penalty)
|
|
costmat[i+1,j+1] = distmat[i,j] + penalty[ipenalty]
|
|
|
|
#enlever les valeurs de l infini
|
|
costmat = costmat[1: , 1:]
|
|
return (costmat, costmat[-1, -1]/(N+M))
|
|
def calculate_mfcc(audio, sr):
|
|
# Define parameters for MFCC calculation
|
|
n_mfcc = 13
|
|
n_fft = 2048
|
|
hop_length = 512
|
|
fmin = 0
|
|
fmax = sr/2
|
|
|
|
# Calculate MFCCs
|
|
mfccs = librosa.feature.mfcc(y=audio, sr=sr, n_mfcc=n_mfcc, n_fft=n_fft, hop_length=hop_length, fmin=fmin, fmax=fmax)
|
|
return mfccs.T
|
|
def calculate_dtw_cost(mfccs_query , mfccs_train):
|
|
distmat = dist.cdist(mfccs_query, mfccs_train,"cosine")
|
|
costmat,mincost = dp(distmat)
|
|
return mincost
|
|
def recognize_speech(audio_query, audio_train_list, sr):#sr frequence d echantillonnage
|
|
# Calculate MFCCs for query audio
|
|
mfccs_query = calculate_mfcc(audio_query, sr)
|
|
|
|
# Calculate DTW cost for each audio in training data
|
|
dtw_costs = []
|
|
for audio_train in audio_train_list:
|
|
mfccs_train = calculate_mfcc(audio_train, sr)
|
|
mincost = calculate_dtw_cost(mfccs_query, mfccs_train)
|
|
dtw_costs.append(mincost)
|
|
|
|
# Find index of word with lowest DTW cost
|
|
index = np.argmin(dtw_costs)
|
|
|
|
# Return recognized word
|
|
return index
|
|
|
|
# Example usage
|
|
def get_recognized_word(recognized_word_index):
|
|
|
|
# Define a dictionary to map recognized word indices to actual words
|
|
word_map = {
|
|
"un" : [0,1,2,3,4,5,6],
|
|
"deux" : [7, 8, 9, 10, 11, 12, 13],
|
|
"trois" : [14, 15, 16, 17, 18, 19],
|
|
"quatre" : [20,21, 22, 23, 24, 25, 26],
|
|
"cinq" : [27 ,28, 29, 30, 31, 32],
|
|
"six" : [33 ,34, 35, 36, 37, 38],
|
|
"sept" : [39 , 40, 41, 42, 43, 44],
|
|
"huit" : [45,46, 47, 48, 49, 50, 51],
|
|
"neuf" : [52,53, 54, 55, 56, 57, 58],
|
|
"dix" : [59,60, 61, 62, 63, 64, 65],
|
|
"bien" : [66 ,67, 68, 69, 70, 71, 72],
|
|
"super" : [127,128,129,130, 131, 132, 133],
|
|
"génial" : [87,88, 89, 90, 91, 92, 93],
|
|
"sympa" : [134,135,136,137, 138, 139, 140],
|
|
"propre" : [122, 123, 124, 125, 126],
|
|
"nul" : [115 ,116, 117, 118, 119, 120, 121],
|
|
"ennuyant" : [80 ,81, 82, 83, 84, 85, 86],
|
|
"j'ai beaucoup aimé" : [94 ,95, 96, 97, 98, 99, 100],
|
|
"j'ai trouvé ça génial" : [101 ,102, 103, 104, 105, 106, 107],
|
|
"je n'ai pas aimé" : [108 ,109, 110, 111, 112, 113, 114],
|
|
"c'était drole" : [73,74, 75, 76, 77, 78, 79],
|
|
}
|
|
for word, indices in word_map.items():
|
|
if recognized_word_index in indices:
|
|
return word
|
|
return "Word not recognized"
|
|
def record_audio(filename, duration, sr):
|
|
chunk = 1024
|
|
sample_format = pyaudio.paInt16
|
|
channels = 1
|
|
record_seconds = duration
|
|
filename = f"{filename}.wav"
|
|
|
|
p = pyaudio.PyAudio()
|
|
|
|
stream = p.open(format=sample_format,
|
|
channels=channels,
|
|
rate=sr,
|
|
frames_per_buffer=chunk,
|
|
input=True)
|
|
|
|
frames = []
|
|
|
|
print(f"Enregistrement en cours...")
|
|
|
|
for i in range(0, int(sr / chunk * record_seconds)):
|
|
data = stream.read(chunk)
|
|
frames.append(data)
|
|
|
|
stream.stop_stream()
|
|
stream.close()
|
|
|
|
p.terminate()
|
|
|
|
|
|
print("Enregistrement terminé")
|
|
|
|
wf = wave.open(filename, "wb")
|
|
wf.setnchannels(channels)
|
|
wf.setsampwidth(p.get_sample_size(sample_format))
|
|
wf.setframerate(sr)
|
|
wf.writeframes(b"".join(frames))
|
|
wf.close()
|
|
|
|
print(f"Fichier enregistré sous {filename}")
|
|
def coupe_silence(signal):
|
|
t = 0
|
|
if signal[t] == 0 :
|
|
p = 0
|
|
while signal[t+p] == 0 :
|
|
if p == 88 :
|
|
signal = signal[:t] + signal[t+p:]
|
|
coupe_silence(signal)
|
|
else :
|
|
p = p+1
|
|
"""
|
|
sr = 44100 # fréquence d'échantillonnage
|
|
duration = 2.5 # durée d'enregistrement en secondes
|
|
filename = "audio_query" # nom du fichier à enregistrer
|
|
|
|
record_audio(filename, duration, sr)
|
|
audio_query, sr = librosa.load('C:\\Users\\HP\\audio_query.wav', sr=sr)
|
|
|
|
audio_train_list = [librosa.load('C:\\Users\\HP\\Documents\\cool.wav', sr=sr)[0], librosa.load('C:\\Users\\HP\\Documents\\formidable.wav', sr=sr)[0], librosa.load('C:\\Users\\HP\\Documents\\cest mauvais.wav', sr=sr)[0] , librosa.load('C:\\Users\\HP\\Documents\\un.wav', sr=sr)[0], librosa.load('C:\\Users\\HP\\Documents\\parfait.wav', sr=sr)[0]]
|
|
recognized_word_index = recognize_speech(audio_query, audio_train_list, sr)
|
|
print(f'Recognized word: {recognized_word_index}')
|
|
"""
|
|
sr = 44100 # fréquence d'échantillonnage
|
|
duration = 6 # durée d'enregistrement en secondes
|
|
filename = "audio_query" # nom du fichier à enregistrer
|
|
|
|
record_audio(filename, duration, sr)
|
|
audio_query, sr = librosa.load('C:\\Users\\HP\\audio_query.wav', sr=sr)
|
|
coupe_silence(audio_query)
|
|
audio_train_list = []
|
|
for file in os.listdir('C:\\Users\\HP\\Documents\\Base de données') :
|
|
audio_train_list.append(librosa.load('C:\\Users\\HP\\Documents\\Base de données\\' + file, sr=sr)[0])
|
|
recognized_word_index = recognize_speech(audio_query, audio_train_list, sr)
|
|
recognized_word = get_recognized_word(recognized_word_index)
|
|
print(f'Recognized word: {recognized_word}')
|