Spaces:
Sleeping
Sleeping
token issue fix
Browse files
app.py
CHANGED
|
@@ -106,6 +106,9 @@ def _get_igbo_asr():
|
|
| 106 |
logger.warning("HF_TOKEN not set - Igbo ASR model requires authentication")
|
| 107 |
return None, None
|
| 108 |
|
|
|
|
|
|
|
|
|
|
| 109 |
try:
|
| 110 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 111 |
logger.info("Lazy-loading Igbo ASR model (gated model)...")
|
|
@@ -118,6 +121,7 @@ def _get_igbo_asr():
|
|
| 118 |
except Exception as e:
|
| 119 |
logger.exception(f"Failed to load Igbo ASR model: {e}")
|
| 120 |
igbo_model, igbo_processor = None, None
|
|
|
|
| 121 |
def _run_whisper(model: WhisperForConditionalGeneration, proc: WhisperProcessor, audio_array: np.ndarray) -> str:
|
| 122 |
try:
|
| 123 |
device = next(model.parameters()).device
|
|
@@ -177,16 +181,18 @@ def speech_to_text(audio_data: bytes) -> str:
|
|
| 177 |
audio_array = preprocess_audio_ffmpeg(audio_data)
|
| 178 |
|
| 179 |
# Try Igbo ASR first for better Igbo detection
|
| 180 |
-
|
| 181 |
-
if
|
|
|
|
| 182 |
igbo_text = _run_whisper(igbo_model, igbo_proc, audio_array)
|
| 183 |
if igbo_text and igbo_text.strip():
|
| 184 |
logger.info("Using Igbo ASR result")
|
| 185 |
return igbo_text
|
| 186 |
|
| 187 |
# Fallback to MMS for other languages
|
| 188 |
-
|
| 189 |
-
if
|
|
|
|
| 190 |
mms_text = _run_mms(mms_model, mms_proc, audio_array)
|
| 191 |
if mms_text and mms_text.strip():
|
| 192 |
logger.info("Using MMS ASR result")
|
|
|
|
| 106 |
logger.warning("HF_TOKEN not set - Igbo ASR model requires authentication")
|
| 107 |
return None, None
|
| 108 |
|
| 109 |
+
|
| 110 |
+
hf_token = hf_token.strip()
|
| 111 |
+
|
| 112 |
try:
|
| 113 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 114 |
logger.info("Lazy-loading Igbo ASR model (gated model)...")
|
|
|
|
| 121 |
except Exception as e:
|
| 122 |
logger.exception(f"Failed to load Igbo ASR model: {e}")
|
| 123 |
igbo_model, igbo_processor = None, None
|
| 124 |
+
return None, None
|
| 125 |
def _run_whisper(model: WhisperForConditionalGeneration, proc: WhisperProcessor, audio_array: np.ndarray) -> str:
|
| 126 |
try:
|
| 127 |
device = next(model.parameters()).device
|
|
|
|
| 181 |
audio_array = preprocess_audio_ffmpeg(audio_data)
|
| 182 |
|
| 183 |
# Try Igbo ASR first for better Igbo detection
|
| 184 |
+
igbo_result = _get_igbo_asr()
|
| 185 |
+
if igbo_result[0] is not None and igbo_result[1] is not None:
|
| 186 |
+
igbo_model, igbo_proc = igbo_result
|
| 187 |
igbo_text = _run_whisper(igbo_model, igbo_proc, audio_array)
|
| 188 |
if igbo_text and igbo_text.strip():
|
| 189 |
logger.info("Using Igbo ASR result")
|
| 190 |
return igbo_text
|
| 191 |
|
| 192 |
# Fallback to MMS for other languages
|
| 193 |
+
mms_result = _get_mms()
|
| 194 |
+
if mms_result[0] is not None and mms_result[1] is not None:
|
| 195 |
+
mms_model, mms_proc = mms_result
|
| 196 |
mms_text = _run_mms(mms_model, mms_proc, audio_array)
|
| 197 |
if mms_text and mms_text.strip():
|
| 198 |
logger.info("Using MMS ASR result")
|