Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
@@ -9,21 +9,25 @@ 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 |
HF_TOKEN = os.environ["HF_TOKEN"]
|
|
|
|
|
13 |
device_map = "auto"
|
14 |
|
15 |
-
# 1.
|
16 |
rvq = load_dac_model(tag="latest", model_type="44khz")
|
17 |
rvq.eval()
|
18 |
-
if torch.cuda.is_available():
|
|
|
19 |
|
20 |
-
# 2.
|
21 |
vad_pipe = PyannotePipeline.from_pretrained(
|
22 |
"pyannote/voice-activity-detection",
|
23 |
use_auth_token=HF_TOKEN
|
24 |
)
|
25 |
|
26 |
-
# 3. Ultravox
|
27 |
ultravox_pipe = pipeline(
|
28 |
model="fixie-ai/ultravox-v0_4",
|
29 |
trust_remote_code=True,
|
@@ -31,48 +35,55 @@ ultravox_pipe = pipeline(
|
|
31 |
torch_dtype=torch.float16
|
32 |
)
|
33 |
|
34 |
-
# 4. Audio Diffusion
|
35 |
diff_pipe = DiffusionPipeline.from_pretrained(
|
36 |
"teticio/audio-diffusion-instrumental-hiphop-256"
|
37 |
).to("cuda")
|
38 |
|
39 |
-
# 5. Dia TTS
|
40 |
with init_empty_weights():
|
41 |
-
dia = Dia.from_pretrained("nari-labs/Dia-1.6B")
|
42 |
dia = load_checkpoint_and_dispatch(
|
43 |
dia,
|
44 |
"nari-labs/Dia-1.6B",
|
45 |
-
device_map=
|
46 |
dtype=torch.float16
|
47 |
)
|
48 |
-
|
|
|
49 |
def process_audio(audio):
|
50 |
sr, array = audio
|
51 |
array = array.numpy() if torch.is_tensor(array) else array
|
52 |
|
|
|
53 |
_ = vad_pipe(array, sampling_rate=sr)
|
|
|
|
|
54 |
x = torch.tensor(array).unsqueeze(0).to("cuda")
|
55 |
codes = rvq.encode(x)
|
56 |
decoded = rvq.decode(codes).squeeze().cpu().numpy()
|
57 |
|
|
|
58 |
ultra_out = ultravox_pipe({"array": decoded, "sampling_rate": sr})
|
59 |
text = ultra_out.get("text", "")
|
60 |
|
|
|
61 |
pros = diff_pipe(raw_audio=decoded)["audios"][0]
|
62 |
|
|
|
63 |
tts = dia.generate(f"[emotion:neutral] {text}")
|
64 |
tts_np = tts.squeeze().cpu().numpy()
|
65 |
tts_np = tts_np / np.max(np.abs(tts_np)) * 0.95
|
66 |
|
67 |
return (sr, tts_np), text
|
68 |
|
69 |
-
# UI
|
70 |
with gr.Blocks(title="Maya AI 📈") as demo:
|
71 |
gr.Markdown("## Maya-AI: Supernatural Conversational Agent")
|
72 |
-
audio_in
|
73 |
-
send_btn
|
74 |
-
audio_out
|
75 |
-
text_out
|
76 |
send_btn.click(process_audio, inputs=audio_in, outputs=[audio_out, text_out])
|
77 |
|
78 |
if __name__ == "__main__":
|
|
|
9 |
from dac.utils import load_model as load_dac_model
|
10 |
from accelerate import init_empty_weights, load_checkpoint_and_dispatch
|
11 |
|
12 |
+
# Retrieve HF_TOKEN from Secrets
|
13 |
HF_TOKEN = os.environ["HF_TOKEN"]
|
14 |
+
|
15 |
+
# Automatically shard across 4× L4 GPUs
|
16 |
device_map = "auto"
|
17 |
|
18 |
+
# 1. Load Descript Audio Codec (RVQ)
|
19 |
rvq = load_dac_model(tag="latest", model_type="44khz")
|
20 |
rvq.eval()
|
21 |
+
if torch.cuda.is_available():
|
22 |
+
rvq = rvq.to("cuda")
|
23 |
|
24 |
+
# 2. Load Voice Activity Detection via Pyannote
|
25 |
vad_pipe = PyannotePipeline.from_pretrained(
|
26 |
"pyannote/voice-activity-detection",
|
27 |
use_auth_token=HF_TOKEN
|
28 |
)
|
29 |
|
30 |
+
# 3. Load Ultravox (speech-to-text + LLM) via Transformers
|
31 |
ultravox_pipe = pipeline(
|
32 |
model="fixie-ai/ultravox-v0_4",
|
33 |
trust_remote_code=True,
|
|
|
35 |
torch_dtype=torch.float16
|
36 |
)
|
37 |
|
38 |
+
# 4. Load Audio Diffusion model via Diffusers
|
39 |
diff_pipe = DiffusionPipeline.from_pretrained(
|
40 |
"teticio/audio-diffusion-instrumental-hiphop-256"
|
41 |
).to("cuda")
|
42 |
|
43 |
+
# 5. Load Dia TTS with meta-weight initialization and multi-GPU dispatch
|
44 |
with init_empty_weights():
|
45 |
+
dia = Dia.from_pretrained("nari-labs/Dia-1.6B")
|
46 |
dia = load_checkpoint_and_dispatch(
|
47 |
dia,
|
48 |
"nari-labs/Dia-1.6B",
|
49 |
+
device_map=device_map,
|
50 |
dtype=torch.float16
|
51 |
)
|
52 |
+
|
53 |
+
# Inference function
|
54 |
def process_audio(audio):
|
55 |
sr, array = audio
|
56 |
array = array.numpy() if torch.is_tensor(array) else array
|
57 |
|
58 |
+
# 2.1 VAD: segment speech regions (not used further here)
|
59 |
_ = vad_pipe(array, sampling_rate=sr)
|
60 |
+
|
61 |
+
# 1.1 RVQ encode/decode for discrete audio tokens
|
62 |
x = torch.tensor(array).unsqueeze(0).to("cuda")
|
63 |
codes = rvq.encode(x)
|
64 |
decoded = rvq.decode(codes).squeeze().cpu().numpy()
|
65 |
|
66 |
+
# 3. Ultravox ASR + LLM to generate response text
|
67 |
ultra_out = ultravox_pipe({"array": decoded, "sampling_rate": sr})
|
68 |
text = ultra_out.get("text", "")
|
69 |
|
70 |
+
# 4. Diffusion-based prosody enhancement
|
71 |
pros = diff_pipe(raw_audio=decoded)["audios"][0]
|
72 |
|
73 |
+
# 5. Dia TTS synthesis with neutral emotion tag
|
74 |
tts = dia.generate(f"[emotion:neutral] {text}")
|
75 |
tts_np = tts.squeeze().cpu().numpy()
|
76 |
tts_np = tts_np / np.max(np.abs(tts_np)) * 0.95
|
77 |
|
78 |
return (sr, tts_np), text
|
79 |
|
80 |
+
# Gradio UI
|
81 |
with gr.Blocks(title="Maya AI 📈") as demo:
|
82 |
gr.Markdown("## Maya-AI: Supernatural Conversational Agent")
|
83 |
+
audio_in = gr.Audio(source="microphone", type="numpy", label="Your Voice")
|
84 |
+
send_btn = gr.Button("Send")
|
85 |
+
audio_out = gr.Audio(label="AI’s Response")
|
86 |
+
text_out = gr.Textbox(label="Generated Text")
|
87 |
send_btn.click(process_audio, inputs=audio_in, outputs=[audio_out, text_out])
|
88 |
|
89 |
if __name__ == "__main__":
|