Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
@@ -3,84 +3,127 @@ import gradio as gr
|
|
3 |
import torch
|
4 |
import numpy as np
|
5 |
from transformers import pipeline
|
6 |
-
from diffusers import DiffusionPipeline
|
7 |
from pyannote.audio import Pipeline as PyannotePipeline
|
8 |
from dia.model import Dia
|
9 |
from dac.utils import load_model as load_dac_model
|
10 |
-
from accelerate import init_empty_weights, load_checkpoint_and_dispatch
|
11 |
|
12 |
-
|
13 |
-
HF_TOKEN
|
14 |
-
device_map = "auto"
|
15 |
|
16 |
-
|
17 |
-
|
|
|
|
|
|
|
18 |
rvq.eval()
|
19 |
-
if torch.cuda.is_available():
|
|
|
20 |
|
21 |
-
|
|
|
22 |
vad_pipe = PyannotePipeline.from_pretrained(
|
23 |
"pyannote/voice-activity-detection",
|
24 |
use_auth_token=HF_TOKEN
|
25 |
-
)
|
26 |
|
27 |
-
|
|
|
28 |
ultravox_pipe = pipeline(
|
29 |
model="fixie-ai/ultravox-v0_4",
|
30 |
trust_remote_code=True,
|
31 |
device_map=device_map,
|
32 |
torch_dtype=torch.float16
|
33 |
-
)
|
34 |
-
|
35 |
-
#-- 4. Audio Diffusion Model (Prosody)
|
36 |
-
diff_pipe = DiffusionPipeline.from_pretrained(
|
37 |
-
"teticio/audio-diffusion-instrumental-hiphop-256",
|
38 |
-
torch_dtype=torch.float16
|
39 |
-
).to("cuda") # Diffusers-based load[2]
|
40 |
-
|
41 |
-
#-- 5. Dia TTS Model Sharded Across GPUs
|
42 |
-
dia = Dia.from_pretrained(
|
43 |
-
"nari-labs/Dia-1.6B",
|
44 |
-
device_map=device_map,
|
45 |
-
torch_dtype=torch.float16,
|
46 |
-
trust_remote_code=True
|
47 |
-
) # Auto-sharding in Transformers[2]
|
48 |
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
arr = arr.numpy() if torch.is_tensor(arr) else arr
|
53 |
-
|
54 |
-
# VAD segmentation
|
55 |
-
_ = vad_pipe({"waveform": torch.tensor(arr).unsqueeze(0), "sample_rate": sr})
|
56 |
-
|
57 |
-
# RVQ encode/decode
|
58 |
-
x = torch.tensor(arr).unsqueeze(0).to("cuda")
|
59 |
-
codes = rvq.encode(x)
|
60 |
-
decoded = rvq.decode(codes).squeeze().cpu().numpy()
|
61 |
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
|
66 |
-
|
67 |
-
pros = diff_pipe(raw_audio=decoded)["audios"][0]
|
68 |
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
75 |
|
76 |
-
|
77 |
with gr.Blocks(title="Maya AI 📈") as demo:
|
78 |
-
gr.Markdown("
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
84 |
|
85 |
if __name__ == "__main__":
|
86 |
demo.launch()
|
|
|
3 |
import torch
|
4 |
import numpy as np
|
5 |
from transformers import pipeline
|
|
|
6 |
from pyannote.audio import Pipeline as PyannotePipeline
|
7 |
from dia.model import Dia
|
8 |
from dac.utils import load_model as load_dac_model
|
|
|
9 |
|
10 |
+
# Environment setup
|
11 |
+
HF_TOKEN = os.environ["HF_TOKEN"]
|
12 |
+
device_map = "auto"
|
13 |
|
14 |
+
print("Loading models...")
|
15 |
+
|
16 |
+
# 1. Load RVQ Codec
|
17 |
+
print("Loading RVQ Codec...")
|
18 |
+
rvq = load_dac_model(tag="latest", model_type="44khz")
|
19 |
rvq.eval()
|
20 |
+
if torch.cuda.is_available():
|
21 |
+
rvq = rvq.to("cuda")
|
22 |
|
23 |
+
# 2. Load VAD Pipeline
|
24 |
+
print("Loading VAD...")
|
25 |
vad_pipe = PyannotePipeline.from_pretrained(
|
26 |
"pyannote/voice-activity-detection",
|
27 |
use_auth_token=HF_TOKEN
|
28 |
+
)
|
29 |
|
30 |
+
# 3. Load Ultravox Pipeline
|
31 |
+
print("Loading Ultravox...")
|
32 |
ultravox_pipe = pipeline(
|
33 |
model="fixie-ai/ultravox-v0_4",
|
34 |
trust_remote_code=True,
|
35 |
device_map=device_map,
|
36 |
torch_dtype=torch.float16
|
37 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
38 |
|
39 |
+
# 4. Skip Audio Diffusion (causing UNet mismatch)
|
40 |
+
print("Skipping Audio Diffusion due to compatibility issues...")
|
41 |
+
diff_pipe = None
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
42 |
|
43 |
+
# 5. Load Dia TTS (correct method based on current API)
|
44 |
+
print("Loading Dia TTS...")
|
45 |
+
dia = Dia.from_pretrained("nari-labs/Dia-1.6B")
|
46 |
|
47 |
+
print("All models loaded successfully!")
|
|
|
48 |
|
49 |
+
def process_audio(audio):
|
50 |
+
try:
|
51 |
+
if audio is None:
|
52 |
+
return None, "No audio input provided"
|
53 |
+
|
54 |
+
sr, array = audio
|
55 |
+
|
56 |
+
# Ensure numpy array
|
57 |
+
if torch.is_tensor(array):
|
58 |
+
array = array.numpy()
|
59 |
+
|
60 |
+
# VAD processing
|
61 |
+
try:
|
62 |
+
vad_result = vad_pipe({"waveform": torch.tensor(array).unsqueeze(0), "sample_rate": sr})
|
63 |
+
except Exception as e:
|
64 |
+
print(f"VAD processing error: {e}")
|
65 |
+
|
66 |
+
# RVQ encode/decode
|
67 |
+
audio_tensor = torch.tensor(array).unsqueeze(0)
|
68 |
+
if torch.cuda.is_available():
|
69 |
+
audio_tensor = audio_tensor.to("cuda")
|
70 |
+
codes = rvq.encode(audio_tensor)
|
71 |
+
decoded = rvq.decode(codes).squeeze().cpu().numpy()
|
72 |
+
|
73 |
+
# Ultravox ASR + LLM
|
74 |
+
ultra_out = ultravox_pipe({"array": decoded, "sampling_rate": sr})
|
75 |
+
text = ultra_out.get("text", "I understand your audio input.")
|
76 |
+
|
77 |
+
# Skip diffusion processing due to compatibility issues
|
78 |
+
prosody_audio = decoded
|
79 |
+
|
80 |
+
# Dia TTS generation
|
81 |
+
tts_output = dia.generate(f"[emotion:neutral] {text}")
|
82 |
+
|
83 |
+
# Convert to numpy and normalize
|
84 |
+
if torch.is_tensor(tts_output):
|
85 |
+
tts_np = tts_output.squeeze().cpu().numpy()
|
86 |
+
else:
|
87 |
+
tts_np = np.array(tts_output)
|
88 |
+
|
89 |
+
# Normalize audio output
|
90 |
+
if len(tts_np) > 0:
|
91 |
+
tts_np = tts_np / np.max(np.abs(tts_np)) * 0.95
|
92 |
+
|
93 |
+
return (sr, tts_np), text
|
94 |
+
|
95 |
+
except Exception as e:
|
96 |
+
print(f"Error in process_audio: {e}")
|
97 |
+
return None, f"Processing error: {str(e)}"
|
98 |
|
99 |
+
# Gradio Interface
|
100 |
with gr.Blocks(title="Maya AI 📈") as demo:
|
101 |
+
gr.Markdown("# Maya-AI: Supernatural Conversational Agent")
|
102 |
+
gr.Markdown("Record audio to interact with the AI agent that understands emotions and responds naturally.")
|
103 |
+
|
104 |
+
with gr.Row():
|
105 |
+
with gr.Column():
|
106 |
+
audio_in = gr.Audio(
|
107 |
+
source="microphone",
|
108 |
+
type="numpy",
|
109 |
+
label="Record Your Voice"
|
110 |
+
)
|
111 |
+
send_btn = gr.Button("Send", variant="primary")
|
112 |
+
|
113 |
+
with gr.Column():
|
114 |
+
audio_out = gr.Audio(label="AI Response")
|
115 |
+
text_out = gr.Textbox(
|
116 |
+
label="Generated Text",
|
117 |
+
lines=3,
|
118 |
+
placeholder="AI response will appear here..."
|
119 |
+
)
|
120 |
+
|
121 |
+
# Event handler
|
122 |
+
send_btn.click(
|
123 |
+
fn=process_audio,
|
124 |
+
inputs=audio_in,
|
125 |
+
outputs=[audio_out, text_out]
|
126 |
+
)
|
127 |
|
128 |
if __name__ == "__main__":
|
129 |
demo.launch()
|