Update app.py
Browse files
app.py
CHANGED
@@ -2,19 +2,20 @@ import gradio as gr
|
|
2 |
import torch
|
3 |
import numpy as np
|
4 |
import librosa
|
5 |
-
from transformers import pipeline
|
|
|
6 |
|
7 |
-
#
|
8 |
-
# ASR Pipeline (
|
9 |
-
#
|
10 |
asr = pipeline(
|
11 |
"automatic-speech-recognition",
|
12 |
model="facebook/wav2vec2-base-960h"
|
13 |
)
|
14 |
|
15 |
-
#
|
16 |
-
#
|
17 |
-
#
|
18 |
translation_models = {
|
19 |
"Spanish": "Helsinki-NLP/opus-mt-en-es",
|
20 |
"Chinese": "Helsinki-NLP/opus-mt-en-zh",
|
@@ -27,62 +28,143 @@ translation_tasks = {
|
|
27 |
"Japanese": "translation_en_to_ja"
|
28 |
}
|
29 |
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
34 |
}
|
35 |
|
36 |
-
#
|
37 |
-
# Caches
|
38 |
-
#
|
39 |
translator_cache = {}
|
40 |
-
|
41 |
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
|
|
|
|
|
|
47 |
translator = pipeline(task_name, model=model_name)
|
48 |
-
translator_cache[
|
49 |
return translator
|
50 |
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
57 |
try:
|
58 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
59 |
except Exception as e:
|
60 |
-
raise
|
61 |
-
|
62 |
-
|
|
|
63 |
|
64 |
-
|
65 |
-
|
66 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
67 |
def predict(audio, text, target_language):
|
68 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
69 |
if text.strip():
|
70 |
english_text = text.strip()
|
71 |
elif audio is not None:
|
72 |
sample_rate, audio_data = audio
|
|
|
|
|
73 |
if audio_data.dtype not in [np.float32, np.float64]:
|
74 |
audio_data = audio_data.astype(np.float32)
|
|
|
|
|
75 |
if len(audio_data.shape) > 1 and audio_data.shape[1] > 1:
|
76 |
audio_data = np.mean(audio_data, axis=1)
|
|
|
|
|
77 |
if sample_rate != 16000:
|
78 |
audio_data = librosa.resample(audio_data, orig_sr=sample_rate, target_sr=16000)
|
79 |
-
|
80 |
-
|
|
|
81 |
english_text = asr_result["text"]
|
82 |
else:
|
83 |
return "No input provided.", "", None
|
84 |
|
85 |
-
# Step 2:
|
86 |
translator = get_translator(target_language)
|
87 |
try:
|
88 |
translation_result = translator(english_text)
|
@@ -90,38 +172,37 @@ def predict(audio, text, target_language):
|
|
90 |
except Exception as e:
|
91 |
return english_text, f"Translation error: {e}", None
|
92 |
|
93 |
-
# Step 3:
|
94 |
try:
|
95 |
-
|
96 |
-
tts_result = tts_pipeline(translated_text)
|
97 |
-
synthesized_audio = (tts_result["sample_rate"], tts_result["wav"])
|
98 |
except Exception as e:
|
99 |
return english_text, translated_text, f"TTS error: {e}"
|
100 |
|
101 |
-
return english_text, translated_text,
|
102 |
|
103 |
-
#
|
104 |
-
# Gradio Interface
|
105 |
-
#
|
106 |
iface = gr.Interface(
|
107 |
fn=predict,
|
108 |
inputs=[
|
109 |
gr.Audio(type="numpy", label="Record/Upload English Audio (optional)"),
|
110 |
gr.Textbox(lines=4, placeholder="Or enter English text here", label="English Text Input (optional)"),
|
111 |
-
gr.Dropdown(choices=
|
112 |
],
|
113 |
outputs=[
|
114 |
gr.Textbox(label="English Transcription"),
|
115 |
gr.Textbox(label="Translation (Target Language)"),
|
116 |
gr.Audio(label="Synthesized Speech in Target Language")
|
117 |
],
|
118 |
-
title="Multimodal Language Learning Aid",
|
119 |
description=(
|
120 |
-
"This app
|
121 |
-
"1. English
|
122 |
-
"2.
|
123 |
-
"3.
|
124 |
-
"
|
|
|
125 |
),
|
126 |
allow_flagging="never"
|
127 |
)
|
|
|
2 |
import torch
|
3 |
import numpy as np
|
4 |
import librosa
|
5 |
+
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="facebook/wav2vec2-base-960h"
|
14 |
)
|
15 |
|
16 |
+
# -----------------------------------------------
|
17 |
+
# 2. Translation Models (3 languages)
|
18 |
+
# -----------------------------------------------
|
19 |
translation_models = {
|
20 |
"Spanish": "Helsinki-NLP/opus-mt-en-es",
|
21 |
"Chinese": "Helsinki-NLP/opus-mt-en-zh",
|
|
|
28 |
"Japanese": "translation_en_to_ja"
|
29 |
}
|
30 |
|
31 |
+
# -----------------------------------------------
|
32 |
+
# 3. TTS Model Configurations
|
33 |
+
# We'll load them manually (not with pipeline("text-to-speech"))
|
34 |
+
# -----------------------------------------------
|
35 |
+
# - Spanish (MMS TTS, uses VITS architecture)
|
36 |
+
# - Chinese (MMS TTS, uses VITS architecture)
|
37 |
+
# - Japanese (SpeechT5 or a VITS-based model—here we pick a SpeechT5 example)
|
38 |
+
tts_config = {
|
39 |
+
"Spanish": {
|
40 |
+
"model_id": "facebook/mms-tts-spa",
|
41 |
+
"architecture": "vits" # We'll use VitsModel
|
42 |
+
},
|
43 |
+
"Chinese": {
|
44 |
+
"model_id": "facebook/mms-tts-che",
|
45 |
+
"architecture": "vits"
|
46 |
+
},
|
47 |
+
"Japanese": {
|
48 |
+
"model_id": "esnya/japanese_speecht5_tts",
|
49 |
+
"architecture": "speecht5" # We'll treat this differently
|
50 |
+
}
|
51 |
}
|
52 |
|
53 |
+
# -----------------------------------------------
|
54 |
+
# 4. Caches
|
55 |
+
# -----------------------------------------------
|
56 |
translator_cache = {}
|
57 |
+
tts_model_cache = {} # store (model, tokenizer, architecture)
|
58 |
|
59 |
+
# -----------------------------------------------
|
60 |
+
# 5. Translator Helper
|
61 |
+
# -----------------------------------------------
|
62 |
+
def get_translator(lang):
|
63 |
+
if lang in translator_cache:
|
64 |
+
return translator_cache[lang]
|
65 |
+
model_name = translation_models[lang]
|
66 |
+
task_name = translation_tasks[lang]
|
67 |
translator = pipeline(task_name, model=model_name)
|
68 |
+
translator_cache[lang] = translator
|
69 |
return translator
|
70 |
|
71 |
+
# -----------------------------------------------
|
72 |
+
# 6. TTS Helper
|
73 |
+
# -----------------------------------------------
|
74 |
+
def get_tts_model(lang):
|
75 |
+
"""
|
76 |
+
Loads (model, tokenizer, architecture) from Hugging Face once, then caches.
|
77 |
+
"""
|
78 |
+
if lang in tts_model_cache:
|
79 |
+
return tts_model_cache[lang]
|
80 |
+
|
81 |
+
config = tts_config.get(lang)
|
82 |
+
if config is None:
|
83 |
+
raise ValueError(f"No TTS config found for language: {lang}")
|
84 |
+
|
85 |
+
model_id = config["model_id"]
|
86 |
+
arch = config["architecture"]
|
87 |
+
|
88 |
try:
|
89 |
+
if arch == "vits":
|
90 |
+
# Load a VitsModel + tokenizer
|
91 |
+
model = VitsModel.from_pretrained(model_id)
|
92 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
93 |
+
elif arch == "speecht5":
|
94 |
+
# For a SpeechT5 model, we might do something else
|
95 |
+
# e.g., pipeline("text-to-speech", model=...) if it works
|
96 |
+
# or custom loading if it's also a VITS-based approach
|
97 |
+
# We'll attempt a similar pattern:
|
98 |
+
model = VitsModel.from_pretrained(model_id)
|
99 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
100 |
+
else:
|
101 |
+
raise ValueError(f"Unknown TTS architecture: {arch}")
|
102 |
except Exception as e:
|
103 |
+
raise RuntimeError(f"Failed to load TTS model {model_id}: {e}")
|
104 |
+
|
105 |
+
tts_model_cache[lang] = (model, tokenizer, arch)
|
106 |
+
return tts_model_cache[lang]
|
107 |
|
108 |
+
def run_tts_inference(lang, text):
|
109 |
+
"""
|
110 |
+
Generates waveform using the loaded TTS model and tokenizer.
|
111 |
+
Returns (sample_rate, np_array).
|
112 |
+
"""
|
113 |
+
model, tokenizer, arch = get_tts_model(lang)
|
114 |
+
inputs = tokenizer(text, return_tensors="pt")
|
115 |
+
|
116 |
+
with torch.no_grad():
|
117 |
+
output = model(**inputs)
|
118 |
+
|
119 |
+
# VitsModel output is typically `.waveform`
|
120 |
+
if hasattr(output, "waveform"):
|
121 |
+
waveform_tensor = output.waveform
|
122 |
+
else:
|
123 |
+
# Some models might return a different attribute
|
124 |
+
raise RuntimeError("The TTS model output doesn't have 'waveform' attribute.")
|
125 |
+
|
126 |
+
# Convert to numpy array
|
127 |
+
waveform = waveform_tensor.squeeze().cpu().numpy()
|
128 |
+
|
129 |
+
# Typically, MMS TTS uses 16 kHz
|
130 |
+
sample_rate = 16000
|
131 |
+
return (sample_rate, waveform)
|
132 |
+
|
133 |
+
# -----------------------------------------------
|
134 |
+
# 7. Prediction Function
|
135 |
+
# -----------------------------------------------
|
136 |
def predict(audio, text, target_language):
|
137 |
+
"""
|
138 |
+
1. If text is provided, use it directly as English text.
|
139 |
+
Else, if audio is provided, run ASR.
|
140 |
+
2. Translate English -> target_language.
|
141 |
+
3. Run TTS with the correct approach for that language.
|
142 |
+
"""
|
143 |
+
# Step 1: English text
|
144 |
if text.strip():
|
145 |
english_text = text.strip()
|
146 |
elif audio is not None:
|
147 |
sample_rate, audio_data = audio
|
148 |
+
|
149 |
+
# Convert to float32
|
150 |
if audio_data.dtype not in [np.float32, np.float64]:
|
151 |
audio_data = audio_data.astype(np.float32)
|
152 |
+
|
153 |
+
# Mono
|
154 |
if len(audio_data.shape) > 1 and audio_data.shape[1] > 1:
|
155 |
audio_data = np.mean(audio_data, axis=1)
|
156 |
+
|
157 |
+
# Resample to 16k
|
158 |
if sample_rate != 16000:
|
159 |
audio_data = librosa.resample(audio_data, orig_sr=sample_rate, target_sr=16000)
|
160 |
+
|
161 |
+
asr_input = {"array": audio_data, "sampling_rate": 16000}
|
162 |
+
asr_result = asr(asr_input)
|
163 |
english_text = asr_result["text"]
|
164 |
else:
|
165 |
return "No input provided.", "", None
|
166 |
|
167 |
+
# Step 2: Translation
|
168 |
translator = get_translator(target_language)
|
169 |
try:
|
170 |
translation_result = translator(english_text)
|
|
|
172 |
except Exception as e:
|
173 |
return english_text, f"Translation error: {e}", None
|
174 |
|
175 |
+
# Step 3: TTS
|
176 |
try:
|
177 |
+
sample_rate, waveform = run_tts_inference(target_language, translated_text)
|
|
|
|
|
178 |
except Exception as e:
|
179 |
return english_text, translated_text, f"TTS error: {e}"
|
180 |
|
181 |
+
return english_text, translated_text, (sample_rate, waveform)
|
182 |
|
183 |
+
# -----------------------------------------------
|
184 |
+
# 8. Gradio Interface
|
185 |
+
# -----------------------------------------------
|
186 |
iface = gr.Interface(
|
187 |
fn=predict,
|
188 |
inputs=[
|
189 |
gr.Audio(type="numpy", label="Record/Upload English Audio (optional)"),
|
190 |
gr.Textbox(lines=4, placeholder="Or enter English text here", label="English Text Input (optional)"),
|
191 |
+
gr.Dropdown(choices=["Spanish", "Chinese", "Japanese"], value="Spanish", label="Target Language")
|
192 |
],
|
193 |
outputs=[
|
194 |
gr.Textbox(label="English Transcription"),
|
195 |
gr.Textbox(label="Translation (Target Language)"),
|
196 |
gr.Audio(label="Synthesized Speech in Target Language")
|
197 |
],
|
198 |
+
title="Multimodal Language Learning Aid (VITS-based TTS)",
|
199 |
description=(
|
200 |
+
"This app:\n"
|
201 |
+
"1. Transcribes English speech (via ASR) or accepts English text.\n"
|
202 |
+
"2. Translates to Spanish, Chinese, or Japanese.\n"
|
203 |
+
"3. Synthesizes speech with VITS-based or SpeechT5-based models.\n\n"
|
204 |
+
"Note: Some models are experimental and may produce errors or poor quality.\n"
|
205 |
+
"Either upload/record English audio or enter text, then select a target language."
|
206 |
),
|
207 |
allow_flagging="never"
|
208 |
)
|