Update app.py
Browse files
app.py
CHANGED
|
@@ -6,11 +6,11 @@ from transformers import pipeline, VitsModel, AutoTokenizer
|
|
| 6 |
import scipy # if needed for processing
|
| 7 |
|
| 8 |
# ------------------------------------------------------
|
| 9 |
-
# 1. ASR Pipeline (English)
|
| 10 |
# ------------------------------------------------------
|
| 11 |
asr = pipeline(
|
| 12 |
"automatic-speech-recognition",
|
| 13 |
-
model="
|
| 14 |
)
|
| 15 |
|
| 16 |
# ------------------------------------------------------
|
|
@@ -30,17 +30,20 @@ translation_tasks = {
|
|
| 30 |
|
| 31 |
# ------------------------------------------------------
|
| 32 |
# 3. TTS Model Configurations
|
| 33 |
-
#
|
| 34 |
-
#
|
| 35 |
# ------------------------------------------------------
|
| 36 |
tts_config = {
|
| 37 |
"Spanish": {
|
| 38 |
"model_id": "facebook/mms-tts-spa", # MMS Spanish
|
| 39 |
"architecture": "vits"
|
| 40 |
},
|
| 41 |
-
"Chinese":
|
|
|
|
|
|
|
|
|
|
| 42 |
"Japanese": {
|
| 43 |
-
"model_id": "
|
| 44 |
"architecture": "vits"
|
| 45 |
}
|
| 46 |
}
|
|
@@ -69,21 +72,19 @@ def get_translator(lang):
|
|
| 69 |
def get_tts_model(lang):
|
| 70 |
"""
|
| 71 |
Loads (model, tokenizer, architecture) from Hugging Face once, then caches.
|
| 72 |
-
If no config is found (e.g. for Chinese), raises ValueError.
|
| 73 |
"""
|
| 74 |
if lang in tts_model_cache:
|
| 75 |
return tts_model_cache[lang]
|
| 76 |
|
| 77 |
config = tts_config.get(lang)
|
| 78 |
if config is None:
|
| 79 |
-
# No TTS model for this language
|
| 80 |
raise ValueError(f"No TTS config found for language: {lang}")
|
| 81 |
|
| 82 |
model_id = config["model_id"]
|
| 83 |
arch = config["architecture"]
|
| 84 |
|
| 85 |
try:
|
| 86 |
-
#
|
| 87 |
model = VitsModel.from_pretrained(model_id)
|
| 88 |
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
| 89 |
except Exception as e:
|
|
@@ -106,17 +107,14 @@ def run_tts_inference(lang, text):
|
|
| 106 |
with torch.no_grad():
|
| 107 |
output = model(**inputs)
|
| 108 |
|
| 109 |
-
# VitsModel output is typically
|
| 110 |
if hasattr(output, "waveform"):
|
| 111 |
waveform_tensor = output.waveform
|
| 112 |
else:
|
| 113 |
raise RuntimeError("TTS model output does not contain 'waveform'.")
|
| 114 |
|
| 115 |
-
# Convert to numpy
|
| 116 |
waveform = waveform_tensor.squeeze().cpu().numpy()
|
| 117 |
-
|
| 118 |
-
# MMS TTS typically uses 16 kHz
|
| 119 |
-
sample_rate = 16000
|
| 120 |
return (sample_rate, waveform)
|
| 121 |
|
| 122 |
# ------------------------------------------------------
|
|
@@ -124,25 +122,25 @@ def run_tts_inference(lang, text):
|
|
| 124 |
# ------------------------------------------------------
|
| 125 |
def predict(audio, text, target_language):
|
| 126 |
"""
|
| 127 |
-
1. Obtain English text (
|
| 128 |
-
2. Translate English
|
| 129 |
-
3.
|
| 130 |
"""
|
| 131 |
-
# Step 1: English text
|
| 132 |
if text.strip():
|
| 133 |
english_text = text.strip()
|
| 134 |
elif audio is not None:
|
| 135 |
sample_rate, audio_data = audio
|
| 136 |
|
| 137 |
-
#
|
| 138 |
if audio_data.dtype not in [np.float32, np.float64]:
|
| 139 |
audio_data = audio_data.astype(np.float32)
|
| 140 |
|
| 141 |
-
# Convert stereo to mono if
|
| 142 |
if len(audio_data.shape) > 1 and audio_data.shape[1] > 1:
|
| 143 |
audio_data = np.mean(audio_data, axis=1)
|
| 144 |
|
| 145 |
-
# Resample to
|
| 146 |
if sample_rate != 16000:
|
| 147 |
audio_data = librosa.resample(audio_data, orig_sr=sample_rate, target_sr=16000)
|
| 148 |
|
|
@@ -160,11 +158,8 @@ def predict(audio, text, target_language):
|
|
| 160 |
except Exception as e:
|
| 161 |
return english_text, f"Translation error: {e}", None
|
| 162 |
|
| 163 |
-
# Step 3: TTS
|
| 164 |
try:
|
| 165 |
-
if tts_config[target_language] is None:
|
| 166 |
-
# No TTS model for Chinese or not supported
|
| 167 |
-
return english_text, translated_text, None
|
| 168 |
sample_rate, waveform = run_tts_inference(target_language, translated_text)
|
| 169 |
except Exception as e:
|
| 170 |
return english_text, translated_text, f"TTS error: {e}"
|
|
@@ -184,20 +179,17 @@ iface = gr.Interface(
|
|
| 184 |
outputs=[
|
| 185 |
gr.Textbox(label="English Transcription"),
|
| 186 |
gr.Textbox(label="Translation (Target Language)"),
|
| 187 |
-
gr.Audio(label="Synthesized Speech
|
| 188 |
],
|
| 189 |
-
title="Multimodal Language Learning Aid (
|
| 190 |
description=(
|
| 191 |
"This app:\n"
|
| 192 |
-
"1. Transcribes English speech
|
| 193 |
-
"2. Translates to Spanish, Chinese, or Japanese (Helsinki-NLP).\n"
|
| 194 |
-
"3.
|
| 195 |
-
"Note: MMS does NOT currently provide a Mandarin TTS model, so TTS is skipped for Chinese."
|
| 196 |
),
|
| 197 |
allow_flagging="never"
|
| 198 |
)
|
| 199 |
|
| 200 |
if __name__ == "__main__":
|
| 201 |
-
# If running locally, uncomment:
|
| 202 |
-
# iface.launch()
|
| 203 |
iface.launch(server_name="0.0.0.0", server_port=7860)
|
|
|
|
| 6 |
import scipy # if needed for processing
|
| 7 |
|
| 8 |
# ------------------------------------------------------
|
| 9 |
+
# 1. ASR Pipeline (English) using Whisper-small
|
| 10 |
# ------------------------------------------------------
|
| 11 |
asr = pipeline(
|
| 12 |
"automatic-speech-recognition",
|
| 13 |
+
model="openai/whisper-small"
|
| 14 |
)
|
| 15 |
|
| 16 |
# ------------------------------------------------------
|
|
|
|
| 30 |
|
| 31 |
# ------------------------------------------------------
|
| 32 |
# 3. TTS Model Configurations
|
| 33 |
+
# For Spanish, we keep the MMS TTS.
|
| 34 |
+
# For Chinese & Japanese, use myshell-ai/MeloTTS-Chinese.
|
| 35 |
# ------------------------------------------------------
|
| 36 |
tts_config = {
|
| 37 |
"Spanish": {
|
| 38 |
"model_id": "facebook/mms-tts-spa", # MMS Spanish
|
| 39 |
"architecture": "vits"
|
| 40 |
},
|
| 41 |
+
"Chinese": {
|
| 42 |
+
"model_id": "myshell-ai/MeloTTS-Chinese",
|
| 43 |
+
"architecture": "vits"
|
| 44 |
+
},
|
| 45 |
"Japanese": {
|
| 46 |
+
"model_id": "myshell-ai/MeloTTS-Japanese",
|
| 47 |
"architecture": "vits"
|
| 48 |
}
|
| 49 |
}
|
|
|
|
| 72 |
def get_tts_model(lang):
|
| 73 |
"""
|
| 74 |
Loads (model, tokenizer, architecture) from Hugging Face once, then caches.
|
|
|
|
| 75 |
"""
|
| 76 |
if lang in tts_model_cache:
|
| 77 |
return tts_model_cache[lang]
|
| 78 |
|
| 79 |
config = tts_config.get(lang)
|
| 80 |
if config is None:
|
|
|
|
| 81 |
raise ValueError(f"No TTS config found for language: {lang}")
|
| 82 |
|
| 83 |
model_id = config["model_id"]
|
| 84 |
arch = config["architecture"]
|
| 85 |
|
| 86 |
try:
|
| 87 |
+
# Assuming the model follows VITS-based inference
|
| 88 |
model = VitsModel.from_pretrained(model_id)
|
| 89 |
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
| 90 |
except Exception as e:
|
|
|
|
| 107 |
with torch.no_grad():
|
| 108 |
output = model(**inputs)
|
| 109 |
|
| 110 |
+
# VitsModel output is typically provided via .waveform attribute
|
| 111 |
if hasattr(output, "waveform"):
|
| 112 |
waveform_tensor = output.waveform
|
| 113 |
else:
|
| 114 |
raise RuntimeError("TTS model output does not contain 'waveform'.")
|
| 115 |
|
|
|
|
| 116 |
waveform = waveform_tensor.squeeze().cpu().numpy()
|
| 117 |
+
sample_rate = 16000 # Typically used sample rate for these models
|
|
|
|
|
|
|
| 118 |
return (sample_rate, waveform)
|
| 119 |
|
| 120 |
# ------------------------------------------------------
|
|
|
|
| 122 |
# ------------------------------------------------------
|
| 123 |
def predict(audio, text, target_language):
|
| 124 |
"""
|
| 125 |
+
1. Obtain English text (via ASR using Whisper-small or text input).
|
| 126 |
+
2. Translate English text to the target language.
|
| 127 |
+
3. Synthesize speech with the target language TTS model.
|
| 128 |
"""
|
| 129 |
+
# Step 1: Get English text
|
| 130 |
if text.strip():
|
| 131 |
english_text = text.strip()
|
| 132 |
elif audio is not None:
|
| 133 |
sample_rate, audio_data = audio
|
| 134 |
|
| 135 |
+
# Ensure float32 data type
|
| 136 |
if audio_data.dtype not in [np.float32, np.float64]:
|
| 137 |
audio_data = audio_data.astype(np.float32)
|
| 138 |
|
| 139 |
+
# Convert stereo to mono if necessary
|
| 140 |
if len(audio_data.shape) > 1 and audio_data.shape[1] > 1:
|
| 141 |
audio_data = np.mean(audio_data, axis=1)
|
| 142 |
|
| 143 |
+
# Resample to 16kHz if necessary
|
| 144 |
if sample_rate != 16000:
|
| 145 |
audio_data = librosa.resample(audio_data, orig_sr=sample_rate, target_sr=16000)
|
| 146 |
|
|
|
|
| 158 |
except Exception as e:
|
| 159 |
return english_text, f"Translation error: {e}", None
|
| 160 |
|
| 161 |
+
# Step 3: TTS
|
| 162 |
try:
|
|
|
|
|
|
|
|
|
|
| 163 |
sample_rate, waveform = run_tts_inference(target_language, translated_text)
|
| 164 |
except Exception as e:
|
| 165 |
return english_text, translated_text, f"TTS error: {e}"
|
|
|
|
| 179 |
outputs=[
|
| 180 |
gr.Textbox(label="English Transcription"),
|
| 181 |
gr.Textbox(label="Translation (Target Language)"),
|
| 182 |
+
gr.Audio(label="Synthesized Speech")
|
| 183 |
],
|
| 184 |
+
title="Multimodal Language Learning Aid (ASR / TTS)",
|
| 185 |
description=(
|
| 186 |
"This app:\n"
|
| 187 |
+
"1. Transcribes English speech or English text.\n"
|
| 188 |
+
"2. Translates to Spanish, Chinese, or Japanese (using Helsinki-NLP models).\n"
|
| 189 |
+
"3. Provides synthetic speech with TTS models:\n"
|
|
|
|
| 190 |
),
|
| 191 |
allow_flagging="never"
|
| 192 |
)
|
| 193 |
|
| 194 |
if __name__ == "__main__":
|
|
|
|
|
|
|
| 195 |
iface.launch(server_name="0.0.0.0", server_port=7860)
|