ggoknar
commited on
Commit
·
a38b58d
1
Parent(s):
da4b074
stream voice with combined wav at end, optional direct stream
Browse files
app.py
CHANGED
|
@@ -5,6 +5,8 @@ import os
|
|
| 5 |
# By using XTTS you agree to CPML license https://coqui.ai/cpml
|
| 6 |
os.environ["COQUI_TOS_AGREED"] = "1"
|
| 7 |
|
|
|
|
|
|
|
| 8 |
import gradio as gr
|
| 9 |
import numpy as np
|
| 10 |
import torch
|
|
@@ -32,6 +34,9 @@ from TTS.utils.generic_utils import get_user_data_dir
|
|
| 32 |
# Could not make play audio next work seemlesly on current Gradio with autoplay so this is a workaround
|
| 33 |
AUDIO_WAIT_MODIFIER = float(os.environ.get("AUDIO_WAIT_MODIFIER", 1))
|
| 34 |
|
|
|
|
|
|
|
|
|
|
| 35 |
# This will trigger downloading model
|
| 36 |
print("Downloading if not downloaded Coqui XTTS V1")
|
| 37 |
tts = TTS("tts_models/multilingual/multi-dataset/xtts_v1")
|
|
@@ -106,3 +111,452 @@ text_client = InferenceClient(
|
|
| 106 |
"mistralai/Mistral-7B-Instruct-v0.1",
|
| 107 |
timeout=WHISPER_TIMEOUT,
|
| 108 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
# By using XTTS you agree to CPML license https://coqui.ai/cpml
|
| 6 |
os.environ["COQUI_TOS_AGREED"] = "1"
|
| 7 |
|
| 8 |
+
from scipy.io.wavfile import write
|
| 9 |
+
from pydub import AudioSegment
|
| 10 |
import gradio as gr
|
| 11 |
import numpy as np
|
| 12 |
import torch
|
|
|
|
| 34 |
# Could not make play audio next work seemlesly on current Gradio with autoplay so this is a workaround
|
| 35 |
AUDIO_WAIT_MODIFIER = float(os.environ.get("AUDIO_WAIT_MODIFIER", 1))
|
| 36 |
|
| 37 |
+
# if set will try to stream audio while receveng audio chunks, beware that recreating audio each time produces artifacts
|
| 38 |
+
DIRECT_STREAM = int(os.environ.get("DIRECT_STREAM", 0))
|
| 39 |
+
|
| 40 |
# This will trigger downloading model
|
| 41 |
print("Downloading if not downloaded Coqui XTTS V1")
|
| 42 |
tts = TTS("tts_models/multilingual/multi-dataset/xtts_v1")
|
|
|
|
| 111 |
"mistralai/Mistral-7B-Instruct-v0.1",
|
| 112 |
timeout=WHISPER_TIMEOUT,
|
| 113 |
)
|
| 114 |
+
|
| 115 |
+
|
| 116 |
+
###### COQUI TTS FUNCTIONS ######
|
| 117 |
+
def get_latents(speaker_wav):
|
| 118 |
+
# create as function as we can populate here with voice cleanup/filtering
|
| 119 |
+
(
|
| 120 |
+
gpt_cond_latent,
|
| 121 |
+
diffusion_conditioning,
|
| 122 |
+
speaker_embedding,
|
| 123 |
+
) = model.get_conditioning_latents(audio_path=speaker_wav)
|
| 124 |
+
return gpt_cond_latent, diffusion_conditioning, speaker_embedding
|
| 125 |
+
|
| 126 |
+
|
| 127 |
+
def format_prompt(message, history):
|
| 128 |
+
prompt = (
|
| 129 |
+
"<s>[INST]"
|
| 130 |
+
+ system_message
|
| 131 |
+
+ "[/INST] I understand, I am a Mistral chatbot with speech by Coqui team.</s>"
|
| 132 |
+
)
|
| 133 |
+
for user_prompt, bot_response in history:
|
| 134 |
+
prompt += f"[INST] {user_prompt} [/INST]"
|
| 135 |
+
prompt += f" {bot_response}</s> "
|
| 136 |
+
prompt += f"[INST] {message} [/INST]"
|
| 137 |
+
return prompt
|
| 138 |
+
|
| 139 |
+
|
| 140 |
+
def generate(
|
| 141 |
+
prompt,
|
| 142 |
+
history,
|
| 143 |
+
temperature=0.9,
|
| 144 |
+
max_new_tokens=256,
|
| 145 |
+
top_p=0.95,
|
| 146 |
+
repetition_penalty=1.0,
|
| 147 |
+
):
|
| 148 |
+
temperature = float(temperature)
|
| 149 |
+
if temperature < 1e-2:
|
| 150 |
+
temperature = 1e-2
|
| 151 |
+
top_p = float(top_p)
|
| 152 |
+
|
| 153 |
+
generate_kwargs = dict(
|
| 154 |
+
temperature=temperature,
|
| 155 |
+
max_new_tokens=max_new_tokens,
|
| 156 |
+
top_p=top_p,
|
| 157 |
+
repetition_penalty=repetition_penalty,
|
| 158 |
+
do_sample=True,
|
| 159 |
+
seed=42,
|
| 160 |
+
)
|
| 161 |
+
|
| 162 |
+
formatted_prompt = format_prompt(prompt, history)
|
| 163 |
+
|
| 164 |
+
try:
|
| 165 |
+
stream = text_client.text_generation(
|
| 166 |
+
formatted_prompt,
|
| 167 |
+
**generate_kwargs,
|
| 168 |
+
stream=True,
|
| 169 |
+
details=True,
|
| 170 |
+
return_full_text=False,
|
| 171 |
+
)
|
| 172 |
+
output = ""
|
| 173 |
+
for response in stream:
|
| 174 |
+
output += response.token.text
|
| 175 |
+
yield output
|
| 176 |
+
|
| 177 |
+
except Exception as e:
|
| 178 |
+
if "Too Many Requests" in str(e):
|
| 179 |
+
print("ERROR: Too many requests on mistral client")
|
| 180 |
+
gr.Warning("Unfortunately Mistral is unable to process")
|
| 181 |
+
output = "Unfortuanately I am not able to process your request now !"
|
| 182 |
+
else:
|
| 183 |
+
print("Unhandled Exception: ", str(e))
|
| 184 |
+
gr.Warning("Unfortunately Mistral is unable to process")
|
| 185 |
+
output = "I do not know what happened but I could not understand you ."
|
| 186 |
+
|
| 187 |
+
return output
|
| 188 |
+
|
| 189 |
+
|
| 190 |
+
def transcribe(wav_path):
|
| 191 |
+
try:
|
| 192 |
+
# get first element from whisper_jax and strip it to delete begin and end space
|
| 193 |
+
return whisper_client.predict(
|
| 194 |
+
wav_path, # str (filepath or URL to file) in 'inputs' Audio component
|
| 195 |
+
"transcribe", # str in 'Task' Radio component
|
| 196 |
+
False, # return_timestamps=False for whisper-jax https://gist.github.com/sanchit-gandhi/781dd7003c5b201bfe16d28634c8d4cf#file-whisper_jax_endpoint-py
|
| 197 |
+
api_name="/predict",
|
| 198 |
+
)[0].strip()
|
| 199 |
+
except:
|
| 200 |
+
gr.Warning("There was a problem with Whisper endpoint, telling a joke for you.")
|
| 201 |
+
return "There was a problem with my voice, tell me joke"
|
| 202 |
+
|
| 203 |
+
|
| 204 |
+
# Chatbot demo with multimodal input (text, markdown, LaTeX, code blocks, image, audio, & video). Plus shows support for streaming text.
|
| 205 |
+
|
| 206 |
+
|
| 207 |
+
def add_text(history, text):
|
| 208 |
+
history = [] if history is None else history
|
| 209 |
+
history = history + [(text, None)]
|
| 210 |
+
return history, gr.update(value="", interactive=False)
|
| 211 |
+
|
| 212 |
+
|
| 213 |
+
def add_file(history, file):
|
| 214 |
+
history = [] if history is None else history
|
| 215 |
+
|
| 216 |
+
try:
|
| 217 |
+
text = transcribe(file)
|
| 218 |
+
print("Transcribed text:", text)
|
| 219 |
+
except Exception as e:
|
| 220 |
+
print(str(e))
|
| 221 |
+
gr.Warning("There was an issue with transcription, please try writing for now")
|
| 222 |
+
# Apply a null text on error
|
| 223 |
+
text = "Transcription seems failed, please tell me a joke about chickens"
|
| 224 |
+
|
| 225 |
+
history = history + [(text, None)]
|
| 226 |
+
return history, gr.update(value="", interactive=False)
|
| 227 |
+
|
| 228 |
+
|
| 229 |
+
##NOTE: not using this as it yields a chacter each time while we need to feed history to TTS
|
| 230 |
+
def bot(history, system_prompt=""):
|
| 231 |
+
history = [] if history is None else history
|
| 232 |
+
|
| 233 |
+
if system_prompt == "":
|
| 234 |
+
system_prompt = system_message
|
| 235 |
+
|
| 236 |
+
history[-1][1] = ""
|
| 237 |
+
for character in generate(history[-1][0], history[:-1]):
|
| 238 |
+
history[-1][1] = character
|
| 239 |
+
yield history
|
| 240 |
+
|
| 241 |
+
|
| 242 |
+
def get_latents(speaker_wav):
|
| 243 |
+
# Generate speaker embedding and latents for TTS
|
| 244 |
+
(
|
| 245 |
+
gpt_cond_latent,
|
| 246 |
+
diffusion_conditioning,
|
| 247 |
+
speaker_embedding,
|
| 248 |
+
) = model.get_conditioning_latents(audio_path=speaker_wav)
|
| 249 |
+
return gpt_cond_latent, diffusion_conditioning, speaker_embedding
|
| 250 |
+
|
| 251 |
+
|
| 252 |
+
latent_map = {}
|
| 253 |
+
latent_map["Female_Voice"] = get_latents("examples/female.wav")
|
| 254 |
+
|
| 255 |
+
|
| 256 |
+
def get_voice(prompt, language, latent_tuple, suffix="0"):
|
| 257 |
+
gpt_cond_latent, diffusion_conditioning, speaker_embedding = latent_tuple
|
| 258 |
+
# Direct version
|
| 259 |
+
t0 = time.time()
|
| 260 |
+
out = model.inference(
|
| 261 |
+
prompt, language, gpt_cond_latent, speaker_embedding, diffusion_conditioning
|
| 262 |
+
)
|
| 263 |
+
inference_time = time.time() - t0
|
| 264 |
+
print(f"I: Time to generate audio: {round(inference_time*1000)} milliseconds")
|
| 265 |
+
real_time_factor = (time.time() - t0) / out["wav"].shape[-1] * 24000
|
| 266 |
+
print(f"Real-time factor (RTF): {real_time_factor}")
|
| 267 |
+
wav_filename = f"output_{suffix}.wav"
|
| 268 |
+
torchaudio.save(wav_filename, torch.tensor(out["wav"]).unsqueeze(0), 24000)
|
| 269 |
+
return wav_filename
|
| 270 |
+
|
| 271 |
+
|
| 272 |
+
def wave_header_chunk(frame_input=b"", channels=1, sample_width=2, sample_rate=24000):
|
| 273 |
+
# This will create a wave header then append the frame input
|
| 274 |
+
# It should be first on a streaming wav file
|
| 275 |
+
# Other frames better should not have it (else you will hear some artifacts each chunk start)
|
| 276 |
+
wav_buf = io.BytesIO()
|
| 277 |
+
with wave.open(wav_buf, "wb") as vfout:
|
| 278 |
+
vfout.setnchannels(channels)
|
| 279 |
+
vfout.setsampwidth(sample_width)
|
| 280 |
+
vfout.setframerate(sample_rate)
|
| 281 |
+
vfout.writeframes(frame_input)
|
| 282 |
+
|
| 283 |
+
wav_buf.seek(0)
|
| 284 |
+
return wav_buf.read()
|
| 285 |
+
|
| 286 |
+
|
| 287 |
+
def get_voice_streaming(prompt, language, latent_tuple, suffix="0"):
|
| 288 |
+
gpt_cond_latent, diffusion_conditioning, speaker_embedding = latent_tuple
|
| 289 |
+
try:
|
| 290 |
+
t0 = time.time()
|
| 291 |
+
chunks = model.inference_stream(
|
| 292 |
+
prompt,
|
| 293 |
+
language,
|
| 294 |
+
gpt_cond_latent,
|
| 295 |
+
speaker_embedding,
|
| 296 |
+
)
|
| 297 |
+
|
| 298 |
+
first_chunk = True
|
| 299 |
+
for i, chunk in enumerate(chunks):
|
| 300 |
+
if first_chunk:
|
| 301 |
+
first_chunk_time = time.time() - t0
|
| 302 |
+
metrics_text = f"Latency to first audio chunk: {round(first_chunk_time*1000)} milliseconds\n"
|
| 303 |
+
first_chunk = False
|
| 304 |
+
print(f"Received chunk {i} of audio length {chunk.shape[-1]}")
|
| 305 |
+
|
| 306 |
+
# In case output is required to be multiple voice files
|
| 307 |
+
# out_file = f'{char}_{i}.wav'
|
| 308 |
+
# write(out_file, 24000, chunk.detach().cpu().numpy().squeeze())
|
| 309 |
+
# audio = AudioSegment.from_file(out_file)
|
| 310 |
+
# audio.export(out_file, format='wav')
|
| 311 |
+
# return out_file
|
| 312 |
+
# directly return chunk as bytes for streaming
|
| 313 |
+
chunk = chunk.detach().cpu().numpy().squeeze()
|
| 314 |
+
chunk = (chunk * 32767).astype(np.int16)
|
| 315 |
+
|
| 316 |
+
yield chunk.tobytes()
|
| 317 |
+
|
| 318 |
+
except RuntimeError as e:
|
| 319 |
+
if "device-side assert" in str(e):
|
| 320 |
+
# cannot do anything on cuda device side error, need tor estart
|
| 321 |
+
print(
|
| 322 |
+
f"Exit due to: Unrecoverable exception caused by prompt:{sentence}",
|
| 323 |
+
flush=True,
|
| 324 |
+
)
|
| 325 |
+
gr.Warning("Unhandled Exception encounter, please retry in a minute")
|
| 326 |
+
print("Cuda device-assert Runtime encountered need restart")
|
| 327 |
+
|
| 328 |
+
# HF Space specific.. This error is unrecoverable need to restart space
|
| 329 |
+
api.restart_space(repo_id=repo_id)
|
| 330 |
+
else:
|
| 331 |
+
print("RuntimeError: non device-side assert error:", str(e))
|
| 332 |
+
gr.Warning("Unhandled Exception encounter, please retry in a minute")
|
| 333 |
+
return None
|
| 334 |
+
return None
|
| 335 |
+
except:
|
| 336 |
+
return None
|
| 337 |
+
|
| 338 |
+
|
| 339 |
+
def get_sentence(history, system_prompt=""):
|
| 340 |
+
history = [] if history is None else history
|
| 341 |
+
|
| 342 |
+
if system_prompt == "":
|
| 343 |
+
system_prompt = system_message
|
| 344 |
+
|
| 345 |
+
history[-1][1] = ""
|
| 346 |
+
|
| 347 |
+
mistral_start = time.time()
|
| 348 |
+
print("Mistral start")
|
| 349 |
+
sentence_list = []
|
| 350 |
+
sentence_hash_list = []
|
| 351 |
+
|
| 352 |
+
text_to_generate = ""
|
| 353 |
+
for character in generate(history[-1][0], history[:-1]):
|
| 354 |
+
history[-1][1] = character
|
| 355 |
+
# It is coming word by word
|
| 356 |
+
|
| 357 |
+
text_to_generate = nltk.sent_tokenize(history[-1][1].replace("\n", " ").strip())
|
| 358 |
+
|
| 359 |
+
if len(text_to_generate) > 1:
|
| 360 |
+
dif = len(text_to_generate) - len(sentence_list)
|
| 361 |
+
|
| 362 |
+
if dif == 1 and len(sentence_list) != 0:
|
| 363 |
+
continue
|
| 364 |
+
|
| 365 |
+
sentence = text_to_generate[len(sentence_list)]
|
| 366 |
+
# This is expensive replace with hashing!
|
| 367 |
+
sentence_hash = hash(sentence)
|
| 368 |
+
|
| 369 |
+
if sentence_hash not in sentence_hash_list:
|
| 370 |
+
sentence_hash_list.append(sentence_hash)
|
| 371 |
+
sentence_list.append(sentence)
|
| 372 |
+
print("New Sentence: ", sentence)
|
| 373 |
+
yield (sentence, history)
|
| 374 |
+
|
| 375 |
+
# return that final sentence token
|
| 376 |
+
# TODO need a counter that one may be replica as before
|
| 377 |
+
last_sentence = nltk.sent_tokenize(history[-1][1].replace("\n", " ").strip())[-1]
|
| 378 |
+
sentence_hash = hash(last_sentence)
|
| 379 |
+
if sentence_hash not in sentence_hash_list:
|
| 380 |
+
sentence_hash_list.append(sentence_hash)
|
| 381 |
+
sentence_list.append(last_sentence)
|
| 382 |
+
print("New Sentence: ", last_sentence)
|
| 383 |
+
|
| 384 |
+
yield (last_sentence, history)
|
| 385 |
+
|
| 386 |
+
|
| 387 |
+
def generate_speech(history):
|
| 388 |
+
language = "en"
|
| 389 |
+
|
| 390 |
+
wav_bytestream = b""
|
| 391 |
+
for sentence, history in get_sentence(history):
|
| 392 |
+
print(sentence)
|
| 393 |
+
# Sometimes prompt </s> coming on output remove it
|
| 394 |
+
sentence = sentence.replace("</s>", "")
|
| 395 |
+
# A fast fix for last chacter, may produce weird sounds if it is with text
|
| 396 |
+
if sentence[-1] in ["!", "?", ".", ","]:
|
| 397 |
+
# just add a space
|
| 398 |
+
sentence = sentence[:-1] + " " + sentence[-1]
|
| 399 |
+
print("Sentence for speech:", sentence)
|
| 400 |
+
|
| 401 |
+
try:
|
| 402 |
+
# generate speech using precomputed latents
|
| 403 |
+
# This is not streaming but it will be fast
|
| 404 |
+
# wav = get_voice(sentence,language, latent_map["Female_Voice"], suffix=len(wav_list))
|
| 405 |
+
audio_stream = get_voice_streaming(
|
| 406 |
+
sentence, language, latent_map["Female_Voice"]
|
| 407 |
+
)
|
| 408 |
+
wav_chunks = wave_header_chunk()
|
| 409 |
+
frame_length = 0
|
| 410 |
+
for chunk in audio_stream:
|
| 411 |
+
try:
|
| 412 |
+
wav_bytestream += chunk
|
| 413 |
+
if DIRECT_STREAM:
|
| 414 |
+
yield (
|
| 415 |
+
gr.Audio.update(
|
| 416 |
+
value=wave_header_chunk() + chunk, autoplay=True
|
| 417 |
+
),
|
| 418 |
+
history,
|
| 419 |
+
)
|
| 420 |
+
wait_time = len(chunk) / 2 / 24000
|
| 421 |
+
wait_time = AUDIO_WAIT_MODIFIER * wait_time
|
| 422 |
+
print("Sleeping till chunk end")
|
| 423 |
+
time.sleep(wait_time)
|
| 424 |
+
|
| 425 |
+
else:
|
| 426 |
+
wav_chunks += chunk
|
| 427 |
+
frame_length += len(chunk)
|
| 428 |
+
except:
|
| 429 |
+
# hack to continue on playing. sometimes last chunk is empty , will be fixed on next TTS
|
| 430 |
+
continue
|
| 431 |
+
|
| 432 |
+
if not DIRECT_STREAM:
|
| 433 |
+
yield (gr.Audio.update(value=wav_chunks, autoplay=True), history)
|
| 434 |
+
# Streaming wait time calculation
|
| 435 |
+
# audio_length = frame_length / sample_width/ frame_rate
|
| 436 |
+
wait_time = frame_length / 2 / 24000
|
| 437 |
+
|
| 438 |
+
# for non streaming
|
| 439 |
+
# wait_time= librosa.get_duration(path=wav)
|
| 440 |
+
|
| 441 |
+
wait_time = AUDIO_WAIT_MODIFIER * wait_time
|
| 442 |
+
print("Sleeping till audio end")
|
| 443 |
+
time.sleep(wait_time)
|
| 444 |
+
|
| 445 |
+
except RuntimeError as e:
|
| 446 |
+
if "device-side assert" in str(e):
|
| 447 |
+
# cannot do anything on cuda device side error, need tor estart
|
| 448 |
+
print(
|
| 449 |
+
f"Exit due to: Unrecoverable exception caused by prompt:{sentence}",
|
| 450 |
+
flush=True,
|
| 451 |
+
)
|
| 452 |
+
gr.Warning("Unhandled Exception encounter, please retry in a minute")
|
| 453 |
+
print("Cuda device-assert Runtime encountered need restart")
|
| 454 |
+
|
| 455 |
+
# HF Space specific.. This error is unrecoverable need to restart space
|
| 456 |
+
api.restart_space(repo_id=repo_id)
|
| 457 |
+
else:
|
| 458 |
+
print("RuntimeError: non device-side assert error:", str(e))
|
| 459 |
+
raise e
|
| 460 |
+
|
| 461 |
+
# Spoken on autoplay everysencen now produce a concataned one at the one
|
| 462 |
+
# requires pip install ffmpeg-python
|
| 463 |
+
|
| 464 |
+
# files_to_concat= [ffmpeg.input(w) for w in wav_list]
|
| 465 |
+
# combined_file_name="combined.wav"
|
| 466 |
+
# ffmpeg.concat(*files_to_concat,v=0, a=1).output(combined_file_name).run(overwrite_output=True)
|
| 467 |
+
# final_audio.update(value=combined_file_name, visible=True)
|
| 468 |
+
# yield (combined_file_name, history
|
| 469 |
+
|
| 470 |
+
wav_bytestream = wave_header_chunk() + wav_bytestream
|
| 471 |
+
time.sleep(0.3)
|
| 472 |
+
yield (gr.Audio.update(value=None, autoplay=False), history)
|
| 473 |
+
yield (gr.Audio.update(value=wav_bytestream, autoplay=False), history)
|
| 474 |
+
|
| 475 |
+
|
| 476 |
+
css = """
|
| 477 |
+
.bot .chatbot p {
|
| 478 |
+
overflow: hidden; /* Ensures the content is not revealed until the animation */
|
| 479 |
+
//border-right: .15em solid orange; /* The typwriter cursor */
|
| 480 |
+
white-space: nowrap; /* Keeps the content on a single line */
|
| 481 |
+
margin: 0 auto; /* Gives that scrolling effect as the typing happens */
|
| 482 |
+
letter-spacing: .15em; /* Adjust as needed */
|
| 483 |
+
animation:
|
| 484 |
+
typing 3.5s steps(40, end);
|
| 485 |
+
blink-caret .75s step-end infinite;
|
| 486 |
+
}
|
| 487 |
+
|
| 488 |
+
/* The typing effect */
|
| 489 |
+
@keyframes typing {
|
| 490 |
+
from { width: 0 }
|
| 491 |
+
to { width: 100% }
|
| 492 |
+
}
|
| 493 |
+
|
| 494 |
+
/* The typewriter cursor effect */
|
| 495 |
+
@keyframes blink-caret {
|
| 496 |
+
from, to { border-color: transparent }
|
| 497 |
+
50% { border-color: orange; }
|
| 498 |
+
}
|
| 499 |
+
"""
|
| 500 |
+
|
| 501 |
+
with gr.Blocks(title=title) as demo:
|
| 502 |
+
gr.Markdown(DESCRIPTION)
|
| 503 |
+
|
| 504 |
+
chatbot = gr.Chatbot(
|
| 505 |
+
[],
|
| 506 |
+
elem_id="chatbot",
|
| 507 |
+
avatar_images=("examples/lama.jpeg", "examples/lama2.jpeg"),
|
| 508 |
+
bubble_full_width=False,
|
| 509 |
+
)
|
| 510 |
+
|
| 511 |
+
with gr.Row():
|
| 512 |
+
txt = gr.Textbox(
|
| 513 |
+
scale=3,
|
| 514 |
+
show_label=False,
|
| 515 |
+
placeholder="Enter text and press enter, or speak to your microphone",
|
| 516 |
+
container=False,
|
| 517 |
+
)
|
| 518 |
+
txt_btn = gr.Button(value="Submit text", scale=1)
|
| 519 |
+
btn = gr.Audio(source="microphone", type="filepath", scale=4)
|
| 520 |
+
|
| 521 |
+
with gr.Row():
|
| 522 |
+
audio = gr.Audio(
|
| 523 |
+
label="Generated audio response",
|
| 524 |
+
streaming=False,
|
| 525 |
+
autoplay=False,
|
| 526 |
+
interactive=True,
|
| 527 |
+
show_label=True,
|
| 528 |
+
)
|
| 529 |
+
# TODO add a second audio that plays whole sentences (for mobile especially)
|
| 530 |
+
# final_audio = gr.Audio(label="Final audio response", streaming=False, autoplay=False, interactive=False,show_label=True, visible=False)
|
| 531 |
+
|
| 532 |
+
clear_btn = gr.ClearButton([chatbot, audio])
|
| 533 |
+
|
| 534 |
+
txt_msg = txt_btn.click(add_text, [chatbot, txt], [chatbot, txt], queue=False).then(
|
| 535 |
+
generate_speech, chatbot, [audio, chatbot]
|
| 536 |
+
)
|
| 537 |
+
|
| 538 |
+
txt_msg.then(lambda: gr.update(interactive=True), None, [txt], queue=False)
|
| 539 |
+
|
| 540 |
+
txt_msg = txt.submit(add_text, [chatbot, txt], [chatbot, txt], queue=False).then(
|
| 541 |
+
generate_speech, chatbot, [audio, chatbot]
|
| 542 |
+
)
|
| 543 |
+
|
| 544 |
+
txt_msg.then(lambda: gr.update(interactive=True), None, [txt], queue=False)
|
| 545 |
+
|
| 546 |
+
file_msg = btn.stop_recording(
|
| 547 |
+
add_file, [chatbot, btn], [chatbot, txt], queue=False
|
| 548 |
+
).then(generate_speech, chatbot, [audio, chatbot])
|
| 549 |
+
|
| 550 |
+
gr.Markdown(
|
| 551 |
+
"""
|
| 552 |
+
This Space demonstrates how to speak to a chatbot, based solely on open-source models.
|
| 553 |
+
It relies on 3 models:
|
| 554 |
+
1. [Whisper-large-v2](https://huggingface.co/spaces/sanchit-gandhi/whisper-jax) as an ASR model, to transcribe recorded audio to text. It is called through a [gradio client](https://www.gradio.app/docs/client).
|
| 555 |
+
2. [Mistral-7b-instruct](https://huggingface.co/spaces/osanseviero/mistral-super-fast) as the chat model, the actual chat model. It is called from [huggingface_hub](https://huggingface.co/docs/huggingface_hub/guides/inference).
|
| 556 |
+
3. [Coqui's XTTS](https://huggingface.co/spaces/coqui/xtts) as a TTS model, to generate the chatbot answers. This time, the model is hosted locally.
|
| 557 |
+
|
| 558 |
+
Note:
|
| 559 |
+
- By using this demo you agree to the terms of the Coqui Public Model License at https://coqui.ai/cpml"""
|
| 560 |
+
)
|
| 561 |
+
demo.queue()
|
| 562 |
+
demo.launch(debug=True)
|