Spaces:
Build error
Build error
Create speech_conversation_app.py
Browse files- speech_conversation_app.py +325 -0
speech_conversation_app.py
ADDED
@@ -0,0 +1,325 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import time
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
import gradio as gr
|
6 |
+
from transformers import pipeline, AutoModelForCausalLM, AutoTokenizer, AutoProcessor, AutoModelForSpeechSeq2Seq
|
7 |
+
from datasets import load_dataset
|
8 |
+
import soundfile as sf
|
9 |
+
|
10 |
+
# Global variables to track latency
|
11 |
+
latency_ASR = 0.0
|
12 |
+
latency_LLM = 0.0
|
13 |
+
latency_TTS = 0.0
|
14 |
+
|
15 |
+
# Global variables to store conversation state
|
16 |
+
conversation_history = []
|
17 |
+
audio_output = None
|
18 |
+
|
19 |
+
# ASR Models
|
20 |
+
ASR_OPTIONS = {
|
21 |
+
"Whisper Small": "openai/whisper-small",
|
22 |
+
"Wav2Vec2": "facebook/wav2vec2-base-960h"
|
23 |
+
}
|
24 |
+
|
25 |
+
# LLM Models
|
26 |
+
LLM_OPTIONS = {
|
27 |
+
"Llama-2 7B Chat": "meta-llama/Llama-2-7b-chat-hf",
|
28 |
+
"Flan-T5 Small": "google/flan-t5-small"
|
29 |
+
}
|
30 |
+
|
31 |
+
# TTS Models
|
32 |
+
TTS_OPTIONS = {
|
33 |
+
"VITS": "espnet/kan-bayashi_ljspeech_vits",
|
34 |
+
"FastSpeech2": "espnet/kan-bayashi_ljspeech_fastspeech2"
|
35 |
+
}
|
36 |
+
|
37 |
+
# Load models
|
38 |
+
asr_models = {}
|
39 |
+
llm_models = {}
|
40 |
+
tts_models = {}
|
41 |
+
|
42 |
+
def load_asr_model(model_name):
|
43 |
+
"""Load ASR model from Hugging Face"""
|
44 |
+
global asr_models
|
45 |
+
|
46 |
+
if model_name not in asr_models:
|
47 |
+
print(f"Loading ASR model: {model_name}")
|
48 |
+
model_id = ASR_OPTIONS[model_name]
|
49 |
+
|
50 |
+
if "whisper" in model_id:
|
51 |
+
asr_models[model_name] = pipeline("automatic-speech-recognition", model=model_id)
|
52 |
+
else: # wav2vec2
|
53 |
+
processor = AutoProcessor.from_pretrained(model_id)
|
54 |
+
model = AutoModelForSpeechSeq2Seq.from_pretrained(model_id)
|
55 |
+
asr_models[model_name] = {"processor": processor, "model": model}
|
56 |
+
|
57 |
+
return asr_models[model_name]
|
58 |
+
|
59 |
+
def load_llm_model(model_name):
|
60 |
+
"""Load LLM model from Hugging Face"""
|
61 |
+
global llm_models
|
62 |
+
|
63 |
+
if model_name not in llm_models:
|
64 |
+
print(f"Loading LLM model: {model_name}")
|
65 |
+
model_id = LLM_OPTIONS[model_name]
|
66 |
+
|
67 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
68 |
+
model = AutoModelForCausalLM.from_pretrained(
|
69 |
+
model_id,
|
70 |
+
torch_dtype=torch.float16,
|
71 |
+
device_map="auto"
|
72 |
+
)
|
73 |
+
|
74 |
+
llm_models[model_name] = {
|
75 |
+
"model": model,
|
76 |
+
"tokenizer": tokenizer
|
77 |
+
}
|
78 |
+
|
79 |
+
return llm_models[model_name]
|
80 |
+
|
81 |
+
def load_tts_model(model_name):
|
82 |
+
"""Load TTS model using ESPnet"""
|
83 |
+
global tts_models
|
84 |
+
|
85 |
+
if model_name not in tts_models:
|
86 |
+
print(f"Loading TTS model: {model_name}")
|
87 |
+
try:
|
88 |
+
# Import ESPnet TTS modules
|
89 |
+
from espnet2.bin.tts_inference import Text2Speech
|
90 |
+
|
91 |
+
model_id = TTS_OPTIONS[model_name]
|
92 |
+
tts = Text2Speech.from_pretrained(model_id)
|
93 |
+
tts_models[model_name] = tts
|
94 |
+
|
95 |
+
except ImportError:
|
96 |
+
print("ESPnet not installed. Using mock TTS for demonstration.")
|
97 |
+
tts_models[model_name] = "mock_tts"
|
98 |
+
|
99 |
+
return tts_models[model_name]
|
100 |
+
|
101 |
+
def transcribe_audio(audio_data, sr, asr_model_name):
|
102 |
+
"""Transcribe audio using selected ASR model"""
|
103 |
+
global latency_ASR
|
104 |
+
|
105 |
+
start_time = time.time()
|
106 |
+
|
107 |
+
model = load_asr_model(asr_model_name)
|
108 |
+
|
109 |
+
if "whisper" in ASR_OPTIONS[asr_model_name]:
|
110 |
+
result = model({"array": audio_data, "sampling_rate": sr})
|
111 |
+
transcript = result["text"]
|
112 |
+
else: # wav2vec2
|
113 |
+
inputs = model["processor"](audio_data, sampling_rate=sr, return_tensors="pt")
|
114 |
+
with torch.no_grad():
|
115 |
+
outputs = model["model"].generate(**inputs)
|
116 |
+
transcript = model["processor"].batch_decode(outputs, skip_special_tokens=True)[0]
|
117 |
+
|
118 |
+
latency_ASR = time.time() - start_time
|
119 |
+
return transcript
|
120 |
+
|
121 |
+
def generate_response(transcript, llm_model_name, system_prompt):
|
122 |
+
"""Generate response using selected LLM model"""
|
123 |
+
global latency_LLM, conversation_history
|
124 |
+
|
125 |
+
start_time = time.time()
|
126 |
+
|
127 |
+
model_info = load_llm_model(llm_model_name)
|
128 |
+
model = model_info["model"]
|
129 |
+
tokenizer = model_info["tokenizer"]
|
130 |
+
|
131 |
+
# Format the prompt based on the model
|
132 |
+
if "llama" in LLM_OPTIONS[llm_model_name].lower():
|
133 |
+
# Format for Llama models
|
134 |
+
if not conversation_history:
|
135 |
+
conversation_history.append({"role": "system", "content": system_prompt})
|
136 |
+
|
137 |
+
conversation_history.append({"role": "user", "content": transcript})
|
138 |
+
|
139 |
+
prompt = tokenizer.apply_chat_template(
|
140 |
+
conversation_history,
|
141 |
+
tokenize=False,
|
142 |
+
add_generation_prompt=True
|
143 |
+
)
|
144 |
+
else:
|
145 |
+
# Format for T5 models
|
146 |
+
prompt = f"{system_prompt}\nUser: {transcript}\nAssistant:"
|
147 |
+
|
148 |
+
# Generate text
|
149 |
+
input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(model.device)
|
150 |
+
|
151 |
+
with torch.no_grad():
|
152 |
+
outputs = model.generate(
|
153 |
+
input_ids,
|
154 |
+
max_new_tokens=100,
|
155 |
+
temperature=0.7,
|
156 |
+
top_p=0.9,
|
157 |
+
)
|
158 |
+
|
159 |
+
# Decode the response
|
160 |
+
if "llama" in LLM_OPTIONS[llm_model_name].lower():
|
161 |
+
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
162 |
+
# Extract just the assistant's response
|
163 |
+
response = response.split("Assistant: ")[-1].strip()
|
164 |
+
else:
|
165 |
+
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
166 |
+
|
167 |
+
# Add to conversation history
|
168 |
+
conversation_history.append({"role": "assistant", "content": response})
|
169 |
+
|
170 |
+
latency_LLM = time.time() - start_time
|
171 |
+
return response
|
172 |
+
|
173 |
+
def synthesize_speech(text, tts_model_name):
|
174 |
+
"""Synthesize speech using selected TTS model"""
|
175 |
+
global latency_TTS
|
176 |
+
|
177 |
+
start_time = time.time()
|
178 |
+
|
179 |
+
tts = load_tts_model(tts_model_name)
|
180 |
+
|
181 |
+
if tts == "mock_tts":
|
182 |
+
# Mock TTS response for demonstration
|
183 |
+
# In a real implementation, this would use the ESPnet model
|
184 |
+
# Load a sample audio file for demonstration
|
185 |
+
try:
|
186 |
+
sample_rate = 16000
|
187 |
+
# Generate a simple sine wave as demo audio
|
188 |
+
duration = 2 # seconds
|
189 |
+
t = np.linspace(0, duration, int(sample_rate * duration), endpoint=False)
|
190 |
+
audio_data = 0.5 * np.sin(2 * np.pi * 220 * t) # 220 Hz sine wave
|
191 |
+
except Exception as e:
|
192 |
+
print(f"Error generating mock audio: {e}")
|
193 |
+
audio_data = np.zeros(16000) # 1 second of silence
|
194 |
+
sample_rate = 16000
|
195 |
+
else:
|
196 |
+
# Use actual ESPnet TTS model
|
197 |
+
with torch.no_grad():
|
198 |
+
wav = tts(text)["wav"]
|
199 |
+
audio_data = wav.numpy()
|
200 |
+
sample_rate = tts.fs
|
201 |
+
|
202 |
+
latency_TTS = time.time() - start_time
|
203 |
+
return (sample_rate, audio_data)
|
204 |
+
|
205 |
+
def process_speech(
|
206 |
+
audio_input,
|
207 |
+
asr_option,
|
208 |
+
llm_option,
|
209 |
+
tts_option,
|
210 |
+
system_prompt
|
211 |
+
):
|
212 |
+
"""Process speech: ASR -> LLM -> TTS pipeline"""
|
213 |
+
global audio_output
|
214 |
+
|
215 |
+
# Check if audio input is available
|
216 |
+
if audio_input is None:
|
217 |
+
return None, "", "", None
|
218 |
+
|
219 |
+
# Get audio data
|
220 |
+
sr, audio_data = audio_input
|
221 |
+
|
222 |
+
# ASR: Speech to text
|
223 |
+
transcript = transcribe_audio(audio_data, sr, asr_option)
|
224 |
+
|
225 |
+
# LLM: Generate response
|
226 |
+
response = generate_response(transcript, llm_option, system_prompt)
|
227 |
+
|
228 |
+
# TTS: Text to speech
|
229 |
+
audio_output = synthesize_speech(response, tts_option)
|
230 |
+
|
231 |
+
# Return results
|
232 |
+
return audio_input, transcript, response, audio_output
|
233 |
+
|
234 |
+
def display_latency():
|
235 |
+
"""Display latency information"""
|
236 |
+
return f"""
|
237 |
+
ASR Latency: {latency_ASR:.2f} seconds
|
238 |
+
LLM Latency: {latency_LLM:.2f} seconds
|
239 |
+
TTS Latency: {latency_TTS:.2f} seconds
|
240 |
+
Total Latency: {latency_ASR + latency_LLM + latency_TTS:.2f} seconds
|
241 |
+
"""
|
242 |
+
|
243 |
+
def reset_conversation():
|
244 |
+
"""Reset the conversation history"""
|
245 |
+
global conversation_history, audio_output
|
246 |
+
conversation_history = []
|
247 |
+
audio_output = None
|
248 |
+
return None, "", "", None, ""
|
249 |
+
|
250 |
+
# Create Gradio interface
|
251 |
+
with gr.Blocks(title="Conversational Speech System") as demo:
|
252 |
+
gr.Markdown(
|
253 |
+
"""
|
254 |
+
# Conversational Speech System with ASR, LLM, and TTS
|
255 |
+
|
256 |
+
This demo showcases a complete speech-to-speech conversation system using:
|
257 |
+
- **ASR** (Automatic Speech Recognition) to convert your speech to text
|
258 |
+
- **LLM** (Large Language Model) to generate responses
|
259 |
+
- **TTS** (Text-to-Speech) to convert the responses to speech
|
260 |
+
|
261 |
+
Speak into your microphone and the system will respond with synthesized speech.
|
262 |
+
"""
|
263 |
+
)
|
264 |
+
|
265 |
+
with gr.Row():
|
266 |
+
with gr.Column(scale=1):
|
267 |
+
# Input components
|
268 |
+
audio_input = gr.Audio(
|
269 |
+
sources=["microphone"],
|
270 |
+
type="numpy",
|
271 |
+
label="Speak here",
|
272 |
+
)
|
273 |
+
|
274 |
+
system_prompt = gr.Textbox(
|
275 |
+
label="System Prompt (instructions for the LLM)",
|
276 |
+
value="You are a helpful and friendly AI assistant. Keep your responses concise and under 3 sentences."
|
277 |
+
)
|
278 |
+
|
279 |
+
asr_dropdown = gr.Dropdown(
|
280 |
+
choices=list(ASR_OPTIONS.keys()),
|
281 |
+
value=list(ASR_OPTIONS.keys())[0],
|
282 |
+
label="Select ASR Model"
|
283 |
+
)
|
284 |
+
|
285 |
+
llm_dropdown = gr.Dropdown(
|
286 |
+
choices=list(LLM_OPTIONS.keys()),
|
287 |
+
value=list(LLM_OPTIONS.keys())[0],
|
288 |
+
label="Select LLM Model"
|
289 |
+
)
|
290 |
+
|
291 |
+
tts_dropdown = gr.Dropdown(
|
292 |
+
choices=list(TTS_OPTIONS.keys()),
|
293 |
+
value=list(TTS_OPTIONS.keys())[0],
|
294 |
+
label="Select TTS Model"
|
295 |
+
)
|
296 |
+
|
297 |
+
reset_btn = gr.Button("Reset Conversation")
|
298 |
+
|
299 |
+
with gr.Column(scale=1):
|
300 |
+
# Output components
|
301 |
+
user_transcript = gr.Textbox(label="Your Speech (ASR Output)")
|
302 |
+
system_response = gr.Textbox(label="AI Response (LLM Output)")
|
303 |
+
audio_output_component = gr.Audio(label="AI Voice Response", autoplay=True)
|
304 |
+
latency_info = gr.Textbox(label="Performance Metrics")
|
305 |
+
|
306 |
+
# Set up event handlers
|
307 |
+
audio_input.change(
|
308 |
+
process_speech,
|
309 |
+
inputs=[audio_input, asr_dropdown, llm_dropdown, tts_dropdown, system_prompt],
|
310 |
+
outputs=[audio_input, user_transcript, system_response, audio_output_component]
|
311 |
+
).then(
|
312 |
+
display_latency,
|
313 |
+
inputs=[],
|
314 |
+
outputs=[latency_info]
|
315 |
+
)
|
316 |
+
|
317 |
+
reset_btn.click(
|
318 |
+
reset_conversation,
|
319 |
+
inputs=[],
|
320 |
+
outputs=[audio_input, user_transcript, system_response, audio_output_component, latency_info]
|
321 |
+
)
|
322 |
+
|
323 |
+
# Launch the app
|
324 |
+
if __name__ == "__main__":
|
325 |
+
demo.launch()
|