Update app.py
Browse files
app.py
CHANGED
@@ -1,6 +1,6 @@
|
|
1 |
import os
|
2 |
|
3 |
-
# Set cache dirs
|
4 |
os.environ["HF_HOME"] = "/tmp"
|
5 |
os.environ["TRANSFORMERS_CACHE"] = "/tmp"
|
6 |
os.environ["TORCH_HOME"] = "/tmp"
|
@@ -19,7 +19,7 @@ from transformers import VitsModel, AutoTokenizer
|
|
19 |
|
20 |
app = FastAPI()
|
21 |
|
22 |
-
# Load model
|
23 |
model = VitsModel.from_pretrained("Somali-tts/somali_tts_model")
|
24 |
tokenizer = AutoTokenizer.from_pretrained("saleolow/somali-mms-tts")
|
25 |
|
@@ -82,16 +82,31 @@ def normalize_text(text: str) -> str:
|
|
82 |
text = text.replace("ZamZam", "SamSam")
|
83 |
return text
|
84 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
85 |
class TextIn(BaseModel):
|
86 |
inputs: str
|
87 |
|
88 |
@app.post("/synthesize")
|
89 |
-
async def synthesize(data: TextIn, test: bool = Query(False, description="Set true to
|
90 |
if test:
|
91 |
-
# Generate 2-second 440Hz sine wave for testing playback
|
92 |
duration_s = 2.0
|
93 |
sample_rate = 22050
|
94 |
-
t = np.linspace(0, duration_s, int(sample_rate*duration_s), endpoint=False)
|
95 |
freq = 440
|
96 |
waveform = 0.5 * np.sin(2 * math.pi * freq * t).astype(np.float32)
|
97 |
pcm_waveform = (waveform * 32767).astype(np.int16)
|
@@ -101,45 +116,30 @@ async def synthesize(data: TextIn, test: bool = Query(False, description="Set tr
|
|
101 |
buf.seek(0)
|
102 |
|
103 |
print(f"[TEST MODE] Generated test tone: {pcm_waveform.shape[0]} samples, Sample rate: {sample_rate}")
|
104 |
-
|
105 |
return StreamingResponse(buf, media_type="audio/wav")
|
106 |
|
107 |
-
# Normalize input text
|
108 |
text = normalize_text(data.inputs)
|
109 |
-
|
110 |
-
# Tokenize and move to device
|
111 |
inputs = tokenizer(text, return_tensors="pt").to(device)
|
112 |
|
113 |
-
# Generate waveform
|
114 |
with torch.no_grad():
|
115 |
output = model(**inputs)
|
116 |
|
117 |
-
print("
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
print("Processed waveform shape:", waveform.shape)
|
128 |
-
print("Waveform min/max before clip:", waveform.min(), waveform.max())
|
129 |
-
|
130 |
-
waveform = waveform.astype(np.float32)
|
131 |
-
waveform = np.clip(waveform, -1.0, 1.0)
|
132 |
-
|
133 |
-
pcm_waveform = (waveform * 32767).astype(np.int16)
|
134 |
|
135 |
-
print("
|
136 |
-
print("PCM waveform min/max:", pcm_waveform.min(), pcm_waveform.max())
|
137 |
|
138 |
-
buf = io.BytesIO()
|
139 |
sample_rate = getattr(model.config, "sampling_rate", 22050)
|
140 |
print("Sample rate:", sample_rate)
|
141 |
|
142 |
-
|
143 |
-
buf.seek(0)
|
144 |
|
145 |
-
return StreamingResponse(
|
|
|
1 |
import os
|
2 |
|
3 |
+
# Set cache dirs BEFORE imports for permission fix
|
4 |
os.environ["HF_HOME"] = "/tmp"
|
5 |
os.environ["TRANSFORMERS_CACHE"] = "/tmp"
|
6 |
os.environ["TORCH_HOME"] = "/tmp"
|
|
|
19 |
|
20 |
app = FastAPI()
|
21 |
|
22 |
+
# Load model and tokenizer ONCE at startup
|
23 |
model = VitsModel.from_pretrained("Somali-tts/somali_tts_model")
|
24 |
tokenizer = AutoTokenizer.from_pretrained("saleolow/somali-mms-tts")
|
25 |
|
|
|
82 |
text = text.replace("ZamZam", "SamSam")
|
83 |
return text
|
84 |
|
85 |
+
def waveform_to_wav_bytes(waveform: torch.Tensor, sample_rate: int = 22050) -> bytes:
|
86 |
+
np_waveform = waveform.cpu().numpy()
|
87 |
+
|
88 |
+
if np_waveform.ndim == 3:
|
89 |
+
np_waveform = np_waveform[0]
|
90 |
+
if np_waveform.ndim == 2:
|
91 |
+
np_waveform = np_waveform.mean(axis=0)
|
92 |
+
|
93 |
+
np_waveform = np.clip(np_waveform, -1.0, 1.0).astype(np.float32)
|
94 |
+
pcm_waveform = (np_waveform * 32767).astype(np.int16)
|
95 |
+
|
96 |
+
buf = io.BytesIO()
|
97 |
+
scipy.io.wavfile.write(buf, rate=sample_rate, data=pcm_waveform)
|
98 |
+
buf.seek(0)
|
99 |
+
return buf.read()
|
100 |
+
|
101 |
class TextIn(BaseModel):
|
102 |
inputs: str
|
103 |
|
104 |
@app.post("/synthesize")
|
105 |
+
async def synthesize(data: TextIn, test: bool = Query(False, description="Set true to return a test tone")):
|
106 |
if test:
|
|
|
107 |
duration_s = 2.0
|
108 |
sample_rate = 22050
|
109 |
+
t = np.linspace(0, duration_s, int(sample_rate * duration_s), endpoint=False)
|
110 |
freq = 440
|
111 |
waveform = 0.5 * np.sin(2 * math.pi * freq * t).astype(np.float32)
|
112 |
pcm_waveform = (waveform * 32767).astype(np.int16)
|
|
|
116 |
buf.seek(0)
|
117 |
|
118 |
print(f"[TEST MODE] Generated test tone: {pcm_waveform.shape[0]} samples, Sample rate: {sample_rate}")
|
|
|
119 |
return StreamingResponse(buf, media_type="audio/wav")
|
120 |
|
|
|
121 |
text = normalize_text(data.inputs)
|
|
|
|
|
122 |
inputs = tokenizer(text, return_tensors="pt").to(device)
|
123 |
|
|
|
124 |
with torch.no_grad():
|
125 |
output = model(**inputs)
|
126 |
|
127 |
+
print("Model output type:", type(output))
|
128 |
+
# Try to extract waveform safely:
|
129 |
+
if hasattr(output, "waveform"):
|
130 |
+
waveform = output.waveform
|
131 |
+
elif isinstance(output, dict) and "waveform" in output:
|
132 |
+
waveform = output["waveform"]
|
133 |
+
elif isinstance(output, (tuple, list)):
|
134 |
+
waveform = output[0]
|
135 |
+
else:
|
136 |
+
return {"error": "Waveform not found in model output"}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
137 |
|
138 |
+
print("Extracted waveform shape:", waveform.shape)
|
|
|
139 |
|
|
|
140 |
sample_rate = getattr(model.config, "sampling_rate", 22050)
|
141 |
print("Sample rate:", sample_rate)
|
142 |
|
143 |
+
wav_bytes = waveform_to_wav_bytes(waveform, sample_rate=sample_rate)
|
|
|
144 |
|
145 |
+
return StreamingResponse(io.BytesIO(wav_bytes), media_type="audio/wav")
|