Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
@@ -4,17 +4,40 @@ import torch
|
|
4 |
import librosa
|
5 |
from pathlib import Path
|
6 |
import tempfile, torchaudio
|
7 |
-
|
|
|
|
|
8 |
|
9 |
# Load the MARS5 model
|
10 |
mars5, config_class = torch.hub.load('Camb-ai/mars5-tts', 'mars5_english', trust_repo=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
|
12 |
-
|
13 |
-
|
14 |
-
|
|
|
15 |
|
16 |
# Function to process the text and audio input and generate the synthesized output
|
17 |
def synthesize(text, audio_file, transcript):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
18 |
# Load the reference audio
|
19 |
wav, sr = librosa.load(audio_file, sr=mars5.sr, mono=True)
|
20 |
wav = torch.from_numpy(wav)
|
@@ -29,21 +52,52 @@ def synthesize(text, audio_file, transcript):
|
|
29 |
# Save the synthesized audio to a temporary file
|
30 |
output_path = Path(tempfile.mktemp(suffix=".wav"))
|
31 |
torchaudio.save(output_path, wav_out.unsqueeze(0), mars5.sr)
|
32 |
-
|
33 |
return str(output_path)
|
34 |
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
|
48 |
-
|
49 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
4 |
import librosa
|
5 |
from pathlib import Path
|
6 |
import tempfile, torchaudio
|
7 |
+
# from faster_whisper import WhisperModel
|
8 |
+
from transformers import pipeline
|
9 |
+
from uuid import uuid4
|
10 |
|
11 |
# Load the MARS5 model
|
12 |
mars5, config_class = torch.hub.load('Camb-ai/mars5-tts', 'mars5_english', trust_repo=True)
|
13 |
+
# asr_model = WhisperModel("small", device="cpu", compute_type="int8")
|
14 |
+
asr_model = pipeline(
|
15 |
+
"automatic-speech-recognition",
|
16 |
+
model="openai/whisper-medium",
|
17 |
+
chunk_length_s=30,
|
18 |
+
device=torch.device("cuda"),
|
19 |
+
)
|
20 |
|
21 |
+
def transcribe_file(f: str) -> str:
|
22 |
+
predictions = asr_model(f, return_timestamps=True)["chunks"]
|
23 |
+
print(f">>>>>. predictions: {predictions}")
|
24 |
+
return " ".join([prediction["text"] for prediction in predictions])
|
25 |
|
26 |
# Function to process the text and audio input and generate the synthesized output
|
27 |
def synthesize(text, audio_file, transcript):
|
28 |
+
audio_file = Path(audio_file)
|
29 |
+
temp_file = f"{uuid4()}.{audio_file.suffix}"
|
30 |
+
|
31 |
+
# copying the audio_file
|
32 |
+
with open(audio_file, 'rb') as src, open(temp_file, 'wb') as dst:
|
33 |
+
dst.write(src.read())
|
34 |
+
|
35 |
+
audio_file = temp_file
|
36 |
+
|
37 |
+
print(f">>>>> synthesizing! audio_file: {audio_file}")
|
38 |
+
if not transcript:
|
39 |
+
transcript = transcribe_file(audio_file)
|
40 |
+
|
41 |
# Load the reference audio
|
42 |
wav, sr = librosa.load(audio_file, sr=mars5.sr, mono=True)
|
43 |
wav = torch.from_numpy(wav)
|
|
|
52 |
# Save the synthesized audio to a temporary file
|
53 |
output_path = Path(tempfile.mktemp(suffix=".wav"))
|
54 |
torchaudio.save(output_path, wav_out.unsqueeze(0), mars5.sr)
|
|
|
55 |
return str(output_path)
|
56 |
|
57 |
+
defaults = {
|
58 |
+
'temperature': 0.8,
|
59 |
+
'top_k': -1,
|
60 |
+
'top_p': 0.2,
|
61 |
+
'typical_p': 1.0,
|
62 |
+
'freq_penalty': 2.6,
|
63 |
+
'presence_penalty': 0.4,
|
64 |
+
'rep_penalty_window': 100,
|
65 |
+
'max_prompt_phones': 360,
|
66 |
+
'deep_clone': True,
|
67 |
+
'nar_guidance_w': 3
|
68 |
+
}
|
69 |
|
70 |
+
|
71 |
+
with gr.Blocks() as demo:
|
72 |
+
gr.Markdown("## MARS5 TTS Demo\nEnter text and upload an audio file to clone the voice and generate synthesized speech using MARS5 TTS.")
|
73 |
+
text = gr.Textbox(label="Text to synthesize")
|
74 |
+
audio_file = gr.Audio(label="Audio file to clone from", type="filepath")
|
75 |
+
|
76 |
+
generate_btn = gr.Button(label="Generate Synthesized Audio")
|
77 |
+
|
78 |
+
with gr.Accordion("Advanced Settings", open=False):
|
79 |
+
gr.Markdown("additional inference settings\nWARNING: changing these incorrectly may degrade quality.")
|
80 |
+
prompt_text = gr.Textbox(label="Transcript of voice reference")
|
81 |
+
temperature = gr.Slider(minimum=0.01, maximum=3, step=0.01, label="temperature", value=defaults['temperature'])
|
82 |
+
top_k = gr.Slider(minimum=-1, maximum=2000, step=1, label="top_k", value=defaults['top_k'])
|
83 |
+
top_p = gr.Slider(minimum=0.01, maximum=1.0, step=0.01, label="top_p", value=defaults['top_p'])
|
84 |
+
typical_p = gr.Slider(minimum=0.01, maximum=1, step=0.01, label="typical_p", value=defaults['typical_p'])
|
85 |
+
freq_penalty = gr.Slider(minimum=0, maximum=5, step=0.05, label="freq_penalty", value=defaults['freq_penalty'])
|
86 |
+
presence_penalty = gr.Slider(minimum=0, maximum=5, step=0.05, label="presence_penalty", value=defaults['presence_penalty'])
|
87 |
+
rep_penalty_window = gr.Slider(minimum=1, maximum=500, step=1, label="rep_penalty_window", value=defaults['rep_penalty_window'])
|
88 |
+
nar_guidance_w = gr.Slider(minimum=1, maximum=8, step=0.1, label="nar_guidance_w", value=defaults['nar_guidance_w'])
|
89 |
+
meta_n = gr.Slider(minimum=1, maximum=10, step=1, label="meta_n", value=2, interactive=False)
|
90 |
+
deep_clone = gr.Checkbox(value=defaults['deep_clone'], label='deep_clone')
|
91 |
+
|
92 |
+
dummy = gr.Number(label='Example number', visible=False)
|
93 |
+
|
94 |
+
output = gr.Audio(label="Synthesized Audio", type="filepath")
|
95 |
+
def on_click(text, audio_file, prompt_text):
|
96 |
+
print(f">>>> transcript: {prompt_text}; audio_file = {audio_file}")
|
97 |
+
of = synthesize(text, audio_file, prompt_text)
|
98 |
+
print(f">>>> output file: {of}")
|
99 |
+
return of
|
100 |
+
|
101 |
+
generate_btn.click(on_click, inputs=[text, audio_file, prompt_text], outputs=[output])
|
102 |
+
|
103 |
+
demo.launch(share=False)
|