Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -1,28 +1,29 @@
|
|
1 |
-
import
|
2 |
import requests
|
3 |
import Levenshtein
|
4 |
-
|
5 |
-
from
|
6 |
|
7 |
# Function to securely load the Hugging Face API token
|
8 |
-
@st.cache_resource
|
9 |
def load_hf_token():
|
10 |
-
|
|
|
11 |
|
12 |
# Function to query the Hugging Face Inference API
|
13 |
-
def transcribe_audio_hf(
|
14 |
"""
|
15 |
Transcribes speech from an audio file using the Hugging Face Inference API.
|
16 |
Args:
|
17 |
-
|
18 |
Returns:
|
19 |
str: The transcription of the speech in the audio file.
|
20 |
"""
|
21 |
API_URL = "https://api-inference.huggingface.co/models/jonatasgrosman/wav2vec2-large-xlsr-53-arabic"
|
22 |
headers = {"Authorization": f"Bearer {load_hf_token()}"}
|
23 |
-
response = requests.post(API_URL, headers=headers, data=
|
24 |
return response.json().get("text", "").strip()
|
25 |
|
|
|
26 |
def levenshtein_similarity(transcription1, transcription2):
|
27 |
"""
|
28 |
Calculate the Levenshtein similarity between two transcriptions.
|
@@ -36,102 +37,99 @@ def levenshtein_similarity(transcription1, transcription2):
|
|
36 |
max_len = max(len(transcription1), len(transcription2))
|
37 |
return 1 - distance / max_len # Normalize to get similarity score
|
38 |
|
39 |
-
|
|
|
40 |
"""
|
41 |
Compares the similarity between the transcription of an original audio file and a user's audio file.
|
42 |
Args:
|
43 |
-
|
44 |
-
|
45 |
Returns:
|
46 |
tuple: Transcriptions and Levenshtein similarity score.
|
47 |
"""
|
48 |
-
transcription_original = transcribe_audio_hf(
|
49 |
-
transcription_user = transcribe_audio_hf(
|
50 |
-
|
51 |
-
return transcription_original, transcription_user,
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
user_audio = st.file_uploader("Upload User Audio", type=["wav", "mp3"])
|
72 |
-
|
73 |
-
if user_audio:
|
74 |
-
user_audio_bytes = user_audio.read()
|
75 |
-
st.audio(user_audio_bytes, format="audio/wav")
|
76 |
-
|
77 |
-
# Add a button to perform the test
|
78 |
-
if original_audio_bytes and user_audio_bytes:
|
79 |
-
if st.button("Perform Testing"):
|
80 |
-
with st.spinner("Performing transcription and similarity testing..."):
|
81 |
-
transcription_original, transcription_user, similarity_score = evaluate_audio_similarity(original_audio_bytes, user_audio_bytes)
|
82 |
-
|
83 |
-
# Display results
|
84 |
-
st.markdown("---")
|
85 |
-
st.subheader("Transcriptions and Similarity Score")
|
86 |
-
st.write(f"**Original Transcription:** {transcription_original}")
|
87 |
-
st.write(f"**User Transcription:** {transcription_user}")
|
88 |
-
st.write(f"**Levenshtein Similarity Score:** {similarity_score:.2f}")
|
89 |
-
|
90 |
-
if similarity_score > 0.8: # Adjust the threshold as needed
|
91 |
-
st.success("The pronunciation is likely correct based on transcription similarity.")
|
92 |
-
else:
|
93 |
-
st.error("The pronunciation may be incorrect based on transcription similarity.")
|
94 |
-
|
95 |
-
elif input_method == "Record":
|
96 |
-
st.write("Record or Upload Original Audio")
|
97 |
-
st.write("Click the button below to start recording. The button will turn green when recording is active.")
|
98 |
-
|
99 |
-
original_audio_bytes = audio_recorder(key="original_audio_recorder")
|
100 |
-
|
101 |
-
if original_audio_bytes and len(original_audio_bytes) > 0:
|
102 |
-
with st.spinner("Processing original audio..."):
|
103 |
-
st.audio(original_audio_bytes, format="audio/wav")
|
104 |
-
st.success("Original audio recorded successfully!")
|
105 |
-
|
106 |
else:
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
st.audio(user_audio_bytes, format="audio/wav")
|
117 |
-
st.success("User audio recorded successfully!")
|
118 |
else:
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
import requests
|
3 |
import Levenshtein
|
4 |
+
import numpy as np
|
5 |
+
from transformers import pipeline
|
6 |
|
7 |
# Function to securely load the Hugging Face API token
|
|
|
8 |
def load_hf_token():
|
9 |
+
# Replace this with your actual Hugging Face API token
|
10 |
+
return "your_huggingface_api_token"
|
11 |
|
12 |
# Function to query the Hugging Face Inference API
|
13 |
+
def transcribe_audio_hf(audio):
|
14 |
"""
|
15 |
Transcribes speech from an audio file using the Hugging Face Inference API.
|
16 |
Args:
|
17 |
+
audio (numpy.array): Audio data as a numpy array.
|
18 |
Returns:
|
19 |
str: The transcription of the speech in the audio file.
|
20 |
"""
|
21 |
API_URL = "https://api-inference.huggingface.co/models/jonatasgrosman/wav2vec2-large-xlsr-53-arabic"
|
22 |
headers = {"Authorization": f"Bearer {load_hf_token()}"}
|
23 |
+
response = requests.post(API_URL, headers=headers, data=audio.tobytes())
|
24 |
return response.json().get("text", "").strip()
|
25 |
|
26 |
+
# Function to calculate Levenshtein similarity
|
27 |
def levenshtein_similarity(transcription1, transcription2):
|
28 |
"""
|
29 |
Calculate the Levenshtein similarity between two transcriptions.
|
|
|
37 |
max_len = max(len(transcription1), len(transcription2))
|
38 |
return 1 - distance / max_len # Normalize to get similarity score
|
39 |
|
40 |
+
# Function to evaluate audio similarity
|
41 |
+
def evaluate_audio_similarity(original_audio, user_audio):
|
42 |
"""
|
43 |
Compares the similarity between the transcription of an original audio file and a user's audio file.
|
44 |
Args:
|
45 |
+
original_audio (numpy.array): Original audio data.
|
46 |
+
user_audio (numpy.array): User's audio data.
|
47 |
Returns:
|
48 |
tuple: Transcriptions and Levenshtein similarity score.
|
49 |
"""
|
50 |
+
transcription_original = transcribe_audio_hf(original_audio)
|
51 |
+
transcription_user = transcribe_audio_hf(user_audio)
|
52 |
+
similarity_score = levenshtein_similarity(transcription_original, transcription_user)
|
53 |
+
return transcription_original, transcription_user, similarity_score
|
54 |
+
|
55 |
+
# Set up the Whisper ASR model for full-context and streaming ASR
|
56 |
+
whisper_transcriber = pipeline("automatic-speech-recognition", model="openai/whisper-base.en")
|
57 |
+
|
58 |
+
# Full-context ASR function
|
59 |
+
def full_context_asr(audio):
|
60 |
+
sr, y = audio
|
61 |
+
y = y.astype(np.float32)
|
62 |
+
y /= np.max(np.abs(y))
|
63 |
+
return whisper_transcriber({"sampling_rate": sr, "raw": y})["text"]
|
64 |
+
|
65 |
+
# Streaming ASR function
|
66 |
+
def streaming_asr(stream, new_chunk):
|
67 |
+
sr, y = new_chunk
|
68 |
+
y = y.astype(np.float32)
|
69 |
+
y /= np.max(np.abs(y))
|
70 |
+
|
71 |
+
if stream is not None:
|
72 |
+
stream = np.concatenate([stream, y])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
73 |
else:
|
74 |
+
stream = y
|
75 |
+
|
76 |
+
return stream, whisper_transcriber({"sampling_rate": sr, "raw": stream})["text"]
|
77 |
+
|
78 |
+
# Define Gradio interface for full-context ASR
|
79 |
+
def gradio_full_context_interface(audio):
|
80 |
+
if audio is not None:
|
81 |
+
transcription = full_context_asr(audio)
|
82 |
+
return transcription
|
|
|
|
|
83 |
else:
|
84 |
+
return "Please provide an audio file."
|
85 |
+
|
86 |
+
# Define Gradio interface for audio similarity checking
|
87 |
+
def gradio_similarity_interface(original_audio, user_audio):
|
88 |
+
if original_audio is not None and user_audio is not None:
|
89 |
+
transcription_original, transcription_user, similarity_score = evaluate_audio_similarity(original_audio, user_audio)
|
90 |
+
|
91 |
+
result = {
|
92 |
+
"Original Transcription": transcription_original,
|
93 |
+
"User Transcription": transcription_user,
|
94 |
+
"Levenshtein Similarity Score": similarity_score,
|
95 |
+
}
|
96 |
+
|
97 |
+
if similarity_score > 0.8: # Adjust the threshold as needed
|
98 |
+
result["Feedback"] = "The pronunciation is likely correct based on transcription similarity."
|
99 |
+
else:
|
100 |
+
result["Feedback"] = "The pronunciation may be incorrect based on transcription similarity."
|
101 |
+
|
102 |
+
return result
|
103 |
+
else:
|
104 |
+
return "Please provide both original and user audio for comparison."
|
105 |
+
|
106 |
+
# Create Gradio app for full-context ASR
|
107 |
+
full_context_demo = gr.Interface(
|
108 |
+
fn=gradio_full_context_interface,
|
109 |
+
inputs=gr.Audio(source="microphone", type="numpy"),
|
110 |
+
outputs="text",
|
111 |
+
title="Full-Context ASR Demo"
|
112 |
+
)
|
113 |
+
|
114 |
+
# Create Gradio app for streaming ASR
|
115 |
+
streaming_demo = gr.Interface(
|
116 |
+
fn=streaming_asr,
|
117 |
+
inputs=["state", gr.Audio(source="microphone", type="numpy", streaming=True)],
|
118 |
+
outputs=["state", "text"],
|
119 |
+
live=True,
|
120 |
+
title="Streaming ASR Demo"
|
121 |
+
)
|
122 |
+
|
123 |
+
# Create Gradio app for audio similarity checking
|
124 |
+
similarity_demo = gr.Interface(
|
125 |
+
fn=gradio_similarity_interface,
|
126 |
+
inputs=[
|
127 |
+
gr.Audio(source="upload", type="numpy", label="Original Audio"),
|
128 |
+
gr.Audio(source="upload", type="numpy", label="User Audio")
|
129 |
+
],
|
130 |
+
outputs="json",
|
131 |
+
title="Audio Transcription and Similarity Checker"
|
132 |
+
)
|
133 |
+
|
134 |
+
# Launch all three demos
|
135 |
+
gr.TabbedInterface([full_context_demo, streaming_demo, similarity_demo], ["Full-Context ASR", "Streaming ASR", "Similarity Checker"]).launch()
|