Spaces:
Sleeping
Sleeping
File size: 6,420 Bytes
4c01711 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 |
import random
import Phonemize
from Levenshtein import editops
from gradio.components import Audio, Dropdown, Textbox, Image
import gradio as gr
import transcriber
import json
import pandas as pd
import matplotlib.pyplot as plt
from scipy.io import wavfile
from scipy.signal import spectrogram
import numpy as np
engine = transcriber.transcribe_SA(model_path='models/SA',verbose=0)
phonemizer = Phonemize.phonemization()
prompts = np.loadtxt('data/prompts.txt', dtype=str)
Attributes = engine.att_list
df_output = None
def select_prompt():
return random.choice(prompts)
def phonemize_prompt(prompt):
return ' '.join(phonemizer.cmu_phonemize(prompt)).lower()
def diff_fn():
return [('H','+'),('E','-'),('N',None),('\n', None),('F','-'),('Fgo','-'),('M','+')]
def recognizeAudio(audio_file, attributes):
#print(','.join(attributes))
global df_output
output = engine.transcribe(audio_file, attributes= tuple(attributes), phonological_matrix_file='data/p2att_en_us-arpa.csv', human_readable=False)
records = []
d = json.loads(output)
records.append(['Phoneme']+d['Phoneme']['symbols'])
for att in d['Attributes']:
records.append([att['Name']]+att['Pattern'])
df = pd.DataFrame.from_records(records)
df.fillna('', inplace=True)
df_output = df
return df.to_html(header=False, index=False)
#Get error by matching the expected sequence with the recognized one and return the output in a format that can be visualized by the gradio HighlightedText box
def get_error(exp_list, rec_list):
exp_list = list(exp_list)
rec_list = list(rec_list)
vocab = set(exp_list+rec_list)
w2c = dict(zip(vocab,range(len(vocab))))
exp_out = [[a,None] for a in exp_list]
rec_out = [[a,None] for a in rec_list]
exp_enc = ''.join([chr(w2c[c]) for c in exp_list])
rec_enc = ''.join([chr(w2c[c]) for c in rec_list])
for op, exp_i, rec_i in editops(exp_enc, rec_enc):
if op == 'replace':
exp_out[exp_i][1] = 'S'
rec_out[rec_i][1] = 'S'
elif op == 'insert':
rec_out[rec_i][1] = 'I'
elif op == 'delete':
exp_out[exp_i][1] = 'D'
diff_list = [['Expected:\t', None]] + exp_out + [['\n',None]] + [['Recognized:\t', None]] + rec_out
return diff_list
def scale_vector(vector, new_min, new_max):
min_val = min(vector)
max_val = max(vector)
scaled_vector = []
for val in vector:
scaled_val = ((val - min_val) * (new_max - new_min) / (max_val - min_val)) + new_min
scaled_vector.append(scaled_val)
return scaled_vector
def create_spectrogram_with_att(wav_file, att_contour, att):
# Read the WAV file
sampling_rate, data = wavfile.read(wav_file)
# Calculate the spectrogram
f, t, Sxx = spectrogram(data, fs=sampling_rate)
fig, ax = plt.subplots(figsize=(10, 5))
# Plot the spectrogram
ax.pcolormesh(t, f, 10 * np.log10(Sxx), shading='gouraud')
ax.set_ylabel('Frequency (Hz)')
ax.set_xlabel('Time (s)')
ax.set_title(f'Spectrogram with {att} Contour')
ax.set_ylim(0, 8000) # Adjust the frequency range if necessary
# Plot the att contour
time_pitch = np.arange(0, len(att_contour) * 0.02, 0.02) # Assuming pitch_contour is sampled every 20 ms
ax.plot(time_pitch, att_contour, color='blue', label=f'{att} Contour')
ax.legend()
return fig
def plot_contour(audio_file, att):
indx = engine.processor.tokenizer.convert_tokens_to_ids([f'p_{att}'])
att_contour = engine.logits.squeeze()[:,indx]
att_contour = scale_vector(att_contour, 0, 6000)
fig = create_spectrogram_with_att(audio_file, att_contour, att)
return fig
with gr.Blocks() as gui:
with gr.Tab("Main"):
prompt = gr.Textbox(label='Prompt', value=select_prompt)
get_prompt = gr.Button("Get Prompt")
get_prompt.click(fn=select_prompt, outputs=prompt)
prompt_phonemes = gr.Textbox(label="Expected Phonemes", interactive=False)
get_phoneme = gr.Button("Get Phonemes")
get_phoneme.click(fn=phonemize_prompt, inputs=prompt, outputs=prompt_phonemes)
record_audio = gr.Audio(sources=["microphone","upload"], type="filepath")
att_list = gr.Dropdown(label="Select Attributes", choices=Attributes, value=['vowel', 'voiced', 'consonant'] ,multiselect=True)
process = gr.Button("Process Audio")
recognition = gr.HTML(label='Output')
process.click(fn=recognizeAudio, inputs=[record_audio,att_list], outputs=recognition)
with gr.Tab("Assessment"):
assess = gr.Button("Assessment")
diff = []
for i in range(len(Attributes)+1):
diff.append(gr.HighlightedText(
combine_adjacent=False,
show_legend=True,
color_map={"S": "red", "I": "green", "D":"blue"}, visible=False))
def get_assessment(prompt_phonemes):#, recognized_phonemes, recognized_attributes):
outputs = [gr.HighlightedText(visible=False)]*(len(Attributes)+1)
outputs[0] = gr.HighlightedText(label=f"Phoneme Assessment",
value=get_error(prompt_phonemes.split(), df_output.iloc[0].values[1:]),
visible=True)
i = 1
for i,r in df_output.iloc[1:].iterrows():
convert = lambda ph: '-' if f'n_{att}' in engine.p2att_map[ph] else '+'
att = r.iloc[0]
exp_att = [convert(ph) for ph in prompt_phonemes.split()]
rec_att = r.iloc[1:].values
outputs[i] = gr.HighlightedText(label=f"{att} Assessment",
value=get_error(exp_att, rec_att),
visible=True)
i += 1
return outputs
assess.click(fn=get_assessment, inputs= [prompt_phonemes], outputs=diff)
with gr.Tab("Analysis"):
selected_att = gr.Dropdown( Attributes, label="Select an Attribute to plot", value='voiced', interactive=True)
do_plot = gr.Button('Plot')
plot_block = gr.Plot(label='Spectrogram with Attribute Contour')
do_plot.click(plot_contour, inputs=[record_audio,selected_att], outputs=plot_block)
gui.launch()
|