Update app.py
Browse files
app.py
CHANGED
@@ -15,13 +15,9 @@ tts = pipeline("text-to-speech", model="Baghdad99/english_voice_tts")
|
|
15 |
|
16 |
def translate_speech(audio_input):
|
17 |
print(f"Type of audio: {type(audio_input)}, Value of audio: {audio_input}") # Debug line
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
sample_rate, audio_data = audio_input
|
22 |
-
else:
|
23 |
-
# Load the audio file as a floating point time series
|
24 |
-
audio_data, sample_rate = librosa.load(audio_input, sr=None)
|
25 |
|
26 |
# Normalize the audio data to the range [-1, 1]
|
27 |
audio_data_normalized = audio_data / np.iinfo(audio_data.dtype).max
|
@@ -29,24 +25,20 @@ def translate_speech(audio_input):
|
|
29 |
# Convert the normalized audio data to float64
|
30 |
audio_data_float64 = audio_data_normalized.astype(np.float64)
|
31 |
|
32 |
-
#
|
33 |
-
|
34 |
|
35 |
-
|
|
|
36 |
|
37 |
-
#
|
38 |
-
|
39 |
-
transcription = output["text"]
|
40 |
-
else:
|
41 |
-
print("The output does not contain 'text'")
|
42 |
-
return
|
43 |
|
44 |
-
#
|
45 |
-
|
46 |
|
47 |
# Use the translation pipeline to translate the transcription
|
48 |
translated_text = translator(transcription, return_tensors="pt")
|
49 |
-
print(f"Translated text: {translated_text}") # Print the translated text to see what it contains
|
50 |
|
51 |
# Check if the translated text contains 'generated_token_ids'
|
52 |
if 'generated_token_ids' in translated_text[0]:
|
@@ -56,12 +48,8 @@ def translate_speech(audio_input):
|
|
56 |
print("The translated text does not contain 'generated_token_ids'")
|
57 |
return
|
58 |
|
59 |
-
# Print the translated text string
|
60 |
-
print(f"Translated text string: {translated_text_str}")
|
61 |
-
|
62 |
# Use the text-to-speech pipeline to synthesize the translated text
|
63 |
synthesised_speech = tts(translated_text_str)
|
64 |
-
print(f"Synthesised speech: {synthesised_speech}") # Print the synthesised speech to see what it contains
|
65 |
|
66 |
# Check if the synthesised speech contains 'audio'
|
67 |
if 'audio' in synthesised_speech:
|
@@ -73,14 +61,12 @@ def translate_speech(audio_input):
|
|
73 |
# Flatten the audio data
|
74 |
synthesised_speech_data = synthesised_speech_data.flatten()
|
75 |
|
76 |
-
# Print the shape and type of the synthesised speech data
|
77 |
-
print(f"Synthesised speech data type: {type(synthesised_speech_data)}, Synthesised speech data shape: {synthesised_speech_data.shape}")
|
78 |
-
|
79 |
# Scale the audio data to the range of int16 format
|
80 |
synthesised_speech = (synthesised_speech_data * 32767).astype(np.int16)
|
81 |
|
82 |
return 16000, synthesised_speech
|
83 |
|
|
|
84 |
# Define the Gradio interface
|
85 |
iface = gr.Interface(
|
86 |
fn=translate_speech,
|
|
|
15 |
|
16 |
def translate_speech(audio_input):
|
17 |
print(f"Type of audio: {type(audio_input)}, Value of audio: {audio_input}") # Debug line
|
18 |
+
def translate_speech(audio_input):
|
19 |
+
# Load the audio file as a floating point time series
|
20 |
+
audio_data, sample_rate = librosa.load(audio_input, sr=None)
|
|
|
|
|
|
|
|
|
21 |
|
22 |
# Normalize the audio data to the range [-1, 1]
|
23 |
audio_data_normalized = audio_data / np.iinfo(audio_data.dtype).max
|
|
|
25 |
# Convert the normalized audio data to float64
|
26 |
audio_data_float64 = audio_data_normalized.astype(np.float64)
|
27 |
|
28 |
+
# Prepare the input dictionary
|
29 |
+
input_dict = pipe.tokenizer(audio_data_float64, return_tensors="pt", padding=True)
|
30 |
|
31 |
+
# Use the speech recognition model to get the logits
|
32 |
+
logits = pipe.model(input_dict.input_values.to("cuda")).logits
|
33 |
|
34 |
+
# Get the predicted IDs
|
35 |
+
pred_ids = torch.argmax(logits, dim=-1)[0]
|
|
|
|
|
|
|
|
|
36 |
|
37 |
+
# Decode the predicted IDs to get the transcription
|
38 |
+
transcription = pipe.tokenizer.decode(pred_ids)
|
39 |
|
40 |
# Use the translation pipeline to translate the transcription
|
41 |
translated_text = translator(transcription, return_tensors="pt")
|
|
|
42 |
|
43 |
# Check if the translated text contains 'generated_token_ids'
|
44 |
if 'generated_token_ids' in translated_text[0]:
|
|
|
48 |
print("The translated text does not contain 'generated_token_ids'")
|
49 |
return
|
50 |
|
|
|
|
|
|
|
51 |
# Use the text-to-speech pipeline to synthesize the translated text
|
52 |
synthesised_speech = tts(translated_text_str)
|
|
|
53 |
|
54 |
# Check if the synthesised speech contains 'audio'
|
55 |
if 'audio' in synthesised_speech:
|
|
|
61 |
# Flatten the audio data
|
62 |
synthesised_speech_data = synthesised_speech_data.flatten()
|
63 |
|
|
|
|
|
|
|
64 |
# Scale the audio data to the range of int16 format
|
65 |
synthesised_speech = (synthesised_speech_data * 32767).astype(np.int16)
|
66 |
|
67 |
return 16000, synthesised_speech
|
68 |
|
69 |
+
|
70 |
# Define the Gradio interface
|
71 |
iface = gr.Interface(
|
72 |
fn=translate_speech,
|