Spaces:
Running
on
Zero
Running
on
Zero
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,330 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import torch
|
3 |
+
from transformers import (
|
4 |
+
AutoTokenizer,
|
5 |
+
AutoModelForCausalLM,
|
6 |
+
SpeechT5Processor,
|
7 |
+
SpeechT5ForTextToSpeech,
|
8 |
+
SpeechT5HifiGan,
|
9 |
+
WhisperProcessor, # New: For Speech-to-Text
|
10 |
+
WhisperForConditionalGeneration # New: For Speech-to-Text
|
11 |
+
)
|
12 |
+
from datasets import load_dataset # To get a speaker embedding for TTS
|
13 |
+
import os
|
14 |
+
import spaces # Import the spaces library for GPU decorator
|
15 |
+
import tempfile # For creating temporary audio files
|
16 |
+
import soundfile as sf # To save audio files
|
17 |
+
import librosa # New: For loading audio files for transcription
|
18 |
+
|
19 |
+
# --- Configuration for Language Model (LLM) ---
|
20 |
+
HUGGINGFACE_MODEL_ID = "HuggingFaceH4/Qwen2.5-1.5B-Instruct-gkd"
|
21 |
+
TORCH_DTYPE = torch.bfloat16
|
22 |
+
MAX_NEW_TOKENS = 512
|
23 |
+
DO_SAMPLE = True
|
24 |
+
TEMPERATURE = 0.7
|
25 |
+
TOP_K = 50
|
26 |
+
TOP_P = 0.95
|
27 |
+
|
28 |
+
# --- Configuration for Text-to-Speech (TTS) ---
|
29 |
+
TTS_MODEL_ID = "microsoft/speecht5_tts"
|
30 |
+
TTS_VOCODER_ID = "microsoft/speecht5_hifigan"
|
31 |
+
|
32 |
+
# --- Configuration for Speech-to-Text (STT) ---
|
33 |
+
STT_MODEL_ID = "openai/whisper-tiny" # Using a smaller Whisper model for faster inference
|
34 |
+
|
35 |
+
# --- Global variables for models and tokenizers/processors ---
|
36 |
+
tokenizer = None
|
37 |
+
llm_model = None
|
38 |
+
tts_processor = None
|
39 |
+
tts_model = None
|
40 |
+
tts_vocoder = None
|
41 |
+
speaker_embeddings = None
|
42 |
+
whisper_processor = None # New: Global for Whisper processor
|
43 |
+
whisper_model = None # New: Global for Whisper model
|
44 |
+
|
45 |
+
# --- Load All Models Function ---
|
46 |
+
@spaces.GPU # Decorate with @spaces.GPU to signal this function needs GPU access
|
47 |
+
def load_models():
|
48 |
+
"""
|
49 |
+
Loads the language model, tokenizer, TTS models, speaker embeddings,
|
50 |
+
and STT (Whisper) models from Hugging Face Hub.
|
51 |
+
This function will be called once when the Gradio app starts up.
|
52 |
+
"""
|
53 |
+
global tokenizer, llm_model, tts_processor, tts_model, tts_vocoder, speaker_embeddings
|
54 |
+
global whisper_processor, whisper_model
|
55 |
+
|
56 |
+
if (tokenizer is not None and llm_model is not None and tts_model is not None and
|
57 |
+
whisper_processor is not None and whisper_model is not None):
|
58 |
+
print("All models and tokenizers/processors already loaded.")
|
59 |
+
return
|
60 |
+
|
61 |
+
hf_token = os.environ.get("HF_TOKEN")
|
62 |
+
|
63 |
+
# Load Language Model (LLM)
|
64 |
+
print(f"Loading LLM tokenizer from: {HUGGINGFACE_MODEL_ID}")
|
65 |
+
try:
|
66 |
+
tokenizer = AutoTokenizer.from_pretrained(HUGGINGFACE_MODEL_ID, token=hf_token)
|
67 |
+
if tokenizer.pad_token is None:
|
68 |
+
tokenizer.pad_token = tokenizer.eos_token
|
69 |
+
print(f"Set tokenizer.pad_token to tokenizer.eos_token ({tokenizer.pad_token_id})")
|
70 |
+
|
71 |
+
print(f"Loading LLM model from: {HUGGINGFACE_MODEL_ID}...")
|
72 |
+
llm_model = AutoModelForCausalLM.from_pretrained(
|
73 |
+
HUGGINGFACE_MODEL_ID,
|
74 |
+
torch_dtype=TORCH_DTYPE,
|
75 |
+
device_map="auto",
|
76 |
+
token=hf_token
|
77 |
+
)
|
78 |
+
llm_model.eval()
|
79 |
+
print("LLM model loaded successfully.")
|
80 |
+
except Exception as e:
|
81 |
+
print(f"Error loading LLM model or tokenizer: {e}")
|
82 |
+
raise RuntimeError("Failed to load LLM model. Check your model ID/path and internet connection.")
|
83 |
+
|
84 |
+
# Load TTS models
|
85 |
+
print(f"Loading TTS processor, model, and vocoder from: {TTS_MODEL_ID}, {TTS_VOCODER_ID}")
|
86 |
+
try:
|
87 |
+
tts_processor = SpeechT5Processor.from_pretrained(TTS_MODEL_ID, token=hf_token)
|
88 |
+
tts_model = SpeechT5ForTextToSpeech.from_pretrained(TTS_MODEL_ID, token=hf_token)
|
89 |
+
tts_vocoder = SpeechT5HifiGan.from_pretrained(TTS_VOCODER_ID, token=hf_token)
|
90 |
+
|
91 |
+
print("Loading speaker embeddings for TTS...")
|
92 |
+
embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation", token=hf_token)
|
93 |
+
speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
|
94 |
+
|
95 |
+
device = llm_model.device if llm_model else 'cpu'
|
96 |
+
tts_model.to(device)
|
97 |
+
tts_vocoder.to(device)
|
98 |
+
speaker_embeddings = speaker_embeddings.to(device)
|
99 |
+
print(f"TTS models and speaker embeddings loaded successfully to device: {device}.")
|
100 |
+
|
101 |
+
except Exception as e:
|
102 |
+
print(f"Error loading TTS models or speaker embeddings: {e}")
|
103 |
+
tts_processor = None
|
104 |
+
tts_model = None
|
105 |
+
tts_vocoder = None
|
106 |
+
speaker_embeddings = None
|
107 |
+
raise RuntimeError("Failed to load TTS components. Check model IDs and internet connection.")
|
108 |
+
|
109 |
+
# Load STT (Whisper) model
|
110 |
+
print(f"Loading STT (Whisper) processor and model from: {STT_MODEL_ID}")
|
111 |
+
try:
|
112 |
+
whisper_processor = WhisperProcessor.from_pretrained(STT_MODEL_ID, token=hf_token)
|
113 |
+
whisper_model = WhisperForConditionalGeneration.from_pretrained(STT_MODEL_ID, token=hf_token)
|
114 |
+
|
115 |
+
device = llm_model.device if llm_model else 'cpu' # Use the same device as LLM
|
116 |
+
whisper_model.to(device)
|
117 |
+
print(f"STT (Whisper) model loaded successfully to device: {device}.")
|
118 |
+
except Exception as e:
|
119 |
+
print(f"Error loading STT (Whisper) model or processor: {e}")
|
120 |
+
whisper_processor = None
|
121 |
+
whisper_model = None
|
122 |
+
raise RuntimeError("Failed to load STT (Whisper) components. Check model ID and internet connection.")
|
123 |
+
|
124 |
+
|
125 |
+
# --- Generate Response and Audio Function ---
|
126 |
+
@spaces.GPU # Decorate with @spaces.GPU as this function performs GPU-intensive inference
|
127 |
+
def generate_response_and_audio(
|
128 |
+
message: str, # Current user message
|
129 |
+
history: list # Gradio Chatbot history format (list of dictionaries with 'role' and 'content')
|
130 |
+
) -> tuple: # Returns (updated_history, audio_file_path)
|
131 |
+
"""
|
132 |
+
Generates a text response from the loaded LLM and then converts it to audio
|
133 |
+
using the loaded TTS model.
|
134 |
+
"""
|
135 |
+
global tokenizer, llm_model, tts_processor, tts_model, tts_vocoder, speaker_embeddings
|
136 |
+
|
137 |
+
# Initialize all models if not already loaded
|
138 |
+
if tokenizer is None or llm_model is None or tts_model is None:
|
139 |
+
load_models()
|
140 |
+
|
141 |
+
if tokenizer is None or llm_model is None: # Check LLM loading status
|
142 |
+
history.append({"role": "user", "content": message})
|
143 |
+
history.append({"role": "assistant", "content": "Error: Chatbot LLM not loaded. Please check logs."})
|
144 |
+
return history, None
|
145 |
+
|
146 |
+
# --- 1. Generate Text Response (LLM) ---
|
147 |
+
messages = history
|
148 |
+
messages.append({"role": "user", "content": message})
|
149 |
+
|
150 |
+
try:
|
151 |
+
input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
152 |
+
except Exception as e:
|
153 |
+
print(f"Error applying chat template: {e}")
|
154 |
+
input_text = ""
|
155 |
+
for item in history:
|
156 |
+
if item["role"] == "user":
|
157 |
+
input_text += f"User: {item['content']}\n"
|
158 |
+
elif item["role"] == "assistant":
|
159 |
+
input_text += f"Assistant: {item['content']}\n"
|
160 |
+
input_text += f"User: {message}\nAssistant:"
|
161 |
+
|
162 |
+
input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to(llm_model.device)
|
163 |
+
|
164 |
+
with torch.no_grad():
|
165 |
+
output_ids = llm_model.generate(
|
166 |
+
input_ids,
|
167 |
+
max_new_tokens=MAX_NEW_TOKENS,
|
168 |
+
do_sample=DO_SAMPLE,
|
169 |
+
temperature=TEMPERATURE,
|
170 |
+
top_k=TOP_K,
|
171 |
+
top_p=TOP_P,
|
172 |
+
pad_token_id=tokenizer.eos_token_id
|
173 |
+
)
|
174 |
+
|
175 |
+
generated_token_ids = output_ids[0][input_ids.shape[-1]:]
|
176 |
+
generated_text = tokenizer.decode(generated_token_ids, skip_special_tokens=True).strip()
|
177 |
+
|
178 |
+
# --- 2. Generate Audio from Response (TTS) ---
|
179 |
+
audio_path = None
|
180 |
+
if tts_processor and tts_model and tts_vocoder and speaker_embeddings is not None:
|
181 |
+
try:
|
182 |
+
device = llm_model.device if llm_model else 'cpu'
|
183 |
+
tts_model.to(device)
|
184 |
+
tts_vocoder.to(device)
|
185 |
+
speaker_embeddings = speaker_embeddings.to(device)
|
186 |
+
|
187 |
+
tts_inputs = tts_processor(
|
188 |
+
text=generated_text,
|
189 |
+
return_tensors="pt",
|
190 |
+
max_length=550,
|
191 |
+
truncation=True
|
192 |
+
).to(device)
|
193 |
+
|
194 |
+
with torch.no_grad():
|
195 |
+
speech = tts_model.generate_speech(tts_inputs["input_ids"], speaker_embeddings, vocoder=tts_vocoder)
|
196 |
+
|
197 |
+
with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp_file:
|
198 |
+
audio_path = tmp_file.name
|
199 |
+
sf.write(audio_path, speech.cpu().numpy(), samplerate=16000)
|
200 |
+
print(f"Audio saved to: {audio_path}")
|
201 |
+
|
202 |
+
except Exception as e:
|
203 |
+
print(f"Error generating audio: {e}")
|
204 |
+
audio_path = None
|
205 |
+
else:
|
206 |
+
print("TTS components not loaded. Skipping audio generation.")
|
207 |
+
|
208 |
+
# --- 3. Update Chat History ---
|
209 |
+
history.append({"role": "assistant", "content": generated_text})
|
210 |
+
|
211 |
+
return history, audio_path
|
212 |
+
|
213 |
+
|
214 |
+
# --- Transcribe Audio Function (NEW) ---
|
215 |
+
@spaces.GPU # This function also needs GPU access for Whisper inference
|
216 |
+
def transcribe_audio(audio_filepath):
|
217 |
+
"""
|
218 |
+
Transcribes an audio file using the loaded Whisper model.
|
219 |
+
Handles audio files of varying lengths.
|
220 |
+
"""
|
221 |
+
global whisper_processor, whisper_model
|
222 |
+
|
223 |
+
if whisper_processor is None or whisper_model is None:
|
224 |
+
load_models() # Attempt to load if not already loaded
|
225 |
+
|
226 |
+
if whisper_processor is None or whisper_model is None:
|
227 |
+
return "Error: Speech-to-Text model not loaded. Please check logs."
|
228 |
+
|
229 |
+
if audio_filepath is None:
|
230 |
+
return "No audio input provided for transcription."
|
231 |
+
|
232 |
+
print(f"Transcribing audio from: {audio_filepath}")
|
233 |
+
try:
|
234 |
+
# Load audio file and resample to 16kHz (Whisper's required sample rate)
|
235 |
+
audio, sample_rate = librosa.load(audio_filepath, sr=16000)
|
236 |
+
|
237 |
+
# Process audio input for the Whisper model
|
238 |
+
input_features = whisper_processor(
|
239 |
+
audio,
|
240 |
+
sampling_rate=sample_rate,
|
241 |
+
return_tensors="pt"
|
242 |
+
).input_features.to(whisper_model.device)
|
243 |
+
|
244 |
+
# Generate transcription IDs
|
245 |
+
predicted_ids = whisper_model.generate(input_features)
|
246 |
+
|
247 |
+
# Decode the IDs to text
|
248 |
+
transcription = whisper_processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
|
249 |
+
print(f"Transcription: {transcription}")
|
250 |
+
return transcription
|
251 |
+
|
252 |
+
except Exception as e:
|
253 |
+
print(f"Error during transcription: {e}")
|
254 |
+
return f"Transcription failed: {e}"
|
255 |
+
|
256 |
+
|
257 |
+
# --- Gradio Interface ---
|
258 |
+
with gr.Blocks() as demo:
|
259 |
+
gr.Markdown(
|
260 |
+
"""
|
261 |
+
# HuggingFaceH4/Qwen2.5-1.5B-Instruct-gkd chat bot with Voice Input & Output
|
262 |
+
Type your message or speak into the microphone to chat with the model.
|
263 |
+
The chatbot's response will be spoken, and your audio input can be transcribed!
|
264 |
+
"""
|
265 |
+
)
|
266 |
+
|
267 |
+
with gr.Tab("Chat with Voice"):
|
268 |
+
chatbot = gr.Chatbot(label="Conversation", type='messages')
|
269 |
+
with gr.Row():
|
270 |
+
text_input = gr.Textbox(
|
271 |
+
label="Your message",
|
272 |
+
placeholder="Type your message here...",
|
273 |
+
scale=4
|
274 |
+
)
|
275 |
+
submit_button = gr.Button("Send", scale=1)
|
276 |
+
|
277 |
+
audio_output = gr.Audio(
|
278 |
+
label="Listen to Response",
|
279 |
+
autoplay=True,
|
280 |
+
interactive=False
|
281 |
+
)
|
282 |
+
|
283 |
+
submit_button.click(
|
284 |
+
fn=generate_response_and_audio,
|
285 |
+
inputs=[text_input, chatbot],
|
286 |
+
outputs=[chatbot, audio_output],
|
287 |
+
queue=True
|
288 |
+
)
|
289 |
+
text_input.submit(
|
290 |
+
fn=generate_response_and_audio,
|
291 |
+
inputs=[text_input, chatbot],
|
292 |
+
outputs=[chatbot, audio_output],
|
293 |
+
queue=True
|
294 |
+
)
|
295 |
+
|
296 |
+
with gr.Tab("Audio Transcription"):
|
297 |
+
stt_audio_input = gr.Audio(
|
298 |
+
type="filepath",
|
299 |
+
label="Upload Audio or Record from Microphone",
|
300 |
+
# Removed 'microphone=True' and 'source' as they cause TypeError with older Gradio versions
|
301 |
+
format="wav" # Ensure consistent format
|
302 |
+
)
|
303 |
+
transcribe_button = gr.Button("Transcribe Audio")
|
304 |
+
transcribed_text_output = gr.Textbox(
|
305 |
+
label="Transcription",
|
306 |
+
placeholder="Transcription will appear here...",
|
307 |
+
interactive=False
|
308 |
+
)
|
309 |
+
transcribe_button.click(
|
310 |
+
fn=transcribe_audio,
|
311 |
+
inputs=[stt_audio_input],
|
312 |
+
outputs=[transcribed_text_output],
|
313 |
+
queue=True
|
314 |
+
)
|
315 |
+
|
316 |
+
# Clear button for the entire interface
|
317 |
+
def clear_all():
|
318 |
+
return [], "", None, None, "" # Clear chatbot, text_input, audio_output, stt_audio_input, transcribed_text_output
|
319 |
+
clear_button = gr.Button("Clear All")
|
320 |
+
clear_button.click(
|
321 |
+
clear_all,
|
322 |
+
inputs=None,
|
323 |
+
outputs=[chatbot, text_input, audio_output, stt_audio_input, transcribed_text_output]
|
324 |
+
)
|
325 |
+
|
326 |
+
# Load all models when the app starts up
|
327 |
+
load_models()
|
328 |
+
|
329 |
+
# Launch the Gradio app
|
330 |
+
demo.queue().launch()
|