Spaces:
Runtime error
Runtime error
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,835 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from __future__ import annotations
|
2 |
+
import os
|
3 |
+
#download for mecab
|
4 |
+
os.system('python -m unidic download')
|
5 |
+
|
6 |
+
# we need to compile a CUBLAS version
|
7 |
+
# Or get it from https://jllllll.github.io/llama-cpp-python-cuBLAS-wheels/
|
8 |
+
os.system('CMAKE_ARGS="-DLLAMA_CUBLAS=on" pip install llama-cpp-python==0.2.11')
|
9 |
+
|
10 |
+
# By using XTTS you agree to CPML license https://coqui.ai/cpml
|
11 |
+
os.environ["COQUI_TOS_AGREED"] = "1"
|
12 |
+
|
13 |
+
# NOTE: for streaming will require gradio audio streaming fix
|
14 |
+
# pip install --upgrade -y gradio==0.50.2 git+https://github.com/gorkemgoknar/gradio.git@patch-1
|
15 |
+
|
16 |
+
import textwrap
|
17 |
+
from scipy.io.wavfile import write
|
18 |
+
from pydub import AudioSegment
|
19 |
+
import gradio as gr
|
20 |
+
import numpy as np
|
21 |
+
import torch
|
22 |
+
import nltk # we'll use this to split into sentences
|
23 |
+
nltk.download("punkt")
|
24 |
+
|
25 |
+
import noisereduce as nr
|
26 |
+
import subprocess
|
27 |
+
import langid
|
28 |
+
import uuid
|
29 |
+
import emoji
|
30 |
+
import pathlib
|
31 |
+
|
32 |
+
import datetime
|
33 |
+
|
34 |
+
from scipy.io.wavfile import write
|
35 |
+
from pydub import AudioSegment
|
36 |
+
|
37 |
+
import re
|
38 |
+
import io, wave
|
39 |
+
import librosa
|
40 |
+
import torchaudio
|
41 |
+
from TTS.api import TTS
|
42 |
+
from TTS.tts.configs.xtts_config import XttsConfig
|
43 |
+
from TTS.tts.models.xtts import Xtts
|
44 |
+
from TTS.utils.generic_utils import get_user_data_dir
|
45 |
+
|
46 |
+
|
47 |
+
import gradio as gr
|
48 |
+
import os
|
49 |
+
import time
|
50 |
+
|
51 |
+
import gradio as gr
|
52 |
+
from transformers import pipeline
|
53 |
+
import numpy as np
|
54 |
+
|
55 |
+
from gradio_client import Client
|
56 |
+
from huggingface_hub import InferenceClient
|
57 |
+
|
58 |
+
# This will trigger downloading model
|
59 |
+
print("Downloading if not downloaded Coqui XTTS V2")
|
60 |
+
|
61 |
+
from TTS.utils.manage import ModelManager
|
62 |
+
model_name = "tts_models/multilingual/multi-dataset/xtts_v2"
|
63 |
+
ModelManager().download_model(model_name)
|
64 |
+
model_path = os.path.join(get_user_data_dir("tts"), model_name.replace("/", "--"))
|
65 |
+
print("XTTS downloaded")
|
66 |
+
|
67 |
+
|
68 |
+
print("Loading XTTS")
|
69 |
+
config = XttsConfig()
|
70 |
+
config.load_json(os.path.join(model_path, "config.json"))
|
71 |
+
|
72 |
+
model = Xtts.init_from_config(config)
|
73 |
+
model.load_checkpoint(
|
74 |
+
config,
|
75 |
+
checkpoint_path=os.path.join(model_path, "model.pth"),
|
76 |
+
vocab_path=os.path.join(model_path, "vocab.json"),
|
77 |
+
eval=True,
|
78 |
+
use_deepspeed=True,
|
79 |
+
)
|
80 |
+
model.cuda()
|
81 |
+
print("Done loading TTS")
|
82 |
+
|
83 |
+
#####llm_model = os.environ.get("LLM_MODEL", "mistral") # or "zephyr"
|
84 |
+
|
85 |
+
title = "Voice chat"
|
86 |
+
|
87 |
+
from huggingface_hub import HfApi
|
88 |
+
|
89 |
+
HF_TOKEN = os.environ.get("HF_TOKEN")
|
90 |
+
# will use api to restart space on a unrecoverable error
|
91 |
+
api = HfApi(token=HF_TOKEN)
|
92 |
+
|
93 |
+
repo_id = "coqui/voice-chat-with-zephyr"
|
94 |
+
|
95 |
+
|
96 |
+
default_system_message = f"""
|
97 |
+
You are ##LLM_MODEL###, a large language model trained ##LLM_MODEL_PROVIDER###, architecture of you is decoder-based LM. Your voice backend or text to speech TTS backend is provided via Coqui technology. You are right now served on Huggingface spaces.
|
98 |
+
Don't repeat. Answer short, only few words, as if in a talk. You cannot access the internet, but you have vast knowledge.
|
99 |
+
Current date: CURRENT_DATE .
|
100 |
+
"""
|
101 |
+
|
102 |
+
system_message = os.environ.get("SYSTEM_MESSAGE", default_system_message)
|
103 |
+
system_message = system_message.replace("CURRENT_DATE", str(datetime.date.today()))
|
104 |
+
|
105 |
+
|
106 |
+
# MISTRAL ONLY
|
107 |
+
default_system_understand_message = (
|
108 |
+
"I understand, I am a ##LLM_MODEL### chatbot with speech by Coqui team."
|
109 |
+
)
|
110 |
+
system_understand_message = os.environ.get(
|
111 |
+
"SYSTEM_UNDERSTAND_MESSAGE", default_system_understand_message
|
112 |
+
)
|
113 |
+
|
114 |
+
print("Mistral system message set as:", default_system_message)
|
115 |
+
WHISPER_TIMEOUT = int(os.environ.get("WHISPER_TIMEOUT", 45))
|
116 |
+
|
117 |
+
whisper_client = Client("https://sanchit-gandhi-whisper-large-v2.hf.space/")
|
118 |
+
|
119 |
+
ROLES = ["AI Assistant","AI Beard The Pirate"]
|
120 |
+
|
121 |
+
ROLE_PROMPTS = {}
|
122 |
+
ROLE_PROMPTS["AI Assistant"]=system_message
|
123 |
+
|
124 |
+
#Pirate scenario
|
125 |
+
character_name= "AI Beard"
|
126 |
+
character_scenario= f"As {character_name} you are a 28 year old man who is a pirate on the ship Invisible AI. You are good friends with Guybrush Threepwood and Murray the Skull. Developers did not get you into Monkey Island games as you wanted huge shares of Big Whoop treasure."
|
127 |
+
pirate_system_message = f"You as {character_name}. {character_scenario} Print out only exactly the words that {character_name} would speak out, do not add anything. Don't repeat. Answer short, only few words, as if in a talk. Craft your response only from the first-person perspective of {character_name} and never as user.Current date: #CURRENT_DATE#".replace("#CURRENT_DATE#", str(datetime.date.today()))
|
128 |
+
|
129 |
+
ROLE_PROMPTS["AI Beard The Pirate"]= pirate_system_message
|
130 |
+
##"You are an AI assistant with Zephyr model by Mistral and Hugging Face and speech from Coqui XTTS . User will you give you a task. Your goal is to complete the task as faithfully as you can. While performing the task think step-by-step and justify your steps, your answers should be clear and short sentences"
|
131 |
+
|
132 |
+
### WILL USE LOCAL MISTRAL OR ZEPHYR OR YI
|
133 |
+
### While zephyr and yi will use half GPU to fit all into 16GB, XTTS will use at most 5GB VRAM
|
134 |
+
|
135 |
+
from huggingface_hub import hf_hub_download
|
136 |
+
print("Downloading LLM")
|
137 |
+
print("Downloading Zephyr 7B beta")
|
138 |
+
# #Zephyr
|
139 |
+
hf_hub_download(repo_id="TheBloke/zephyr-7B-beta-GGUF", local_dir=".", filename="zephyr-7b-beta.Q5_K_M.gguf")
|
140 |
+
zephyr_model_path="./zephyr-7b-beta.Q5_K_M.gguf"
|
141 |
+
|
142 |
+
# print("Downloading Mistral 7B Instruct")
|
143 |
+
# #Mistral
|
144 |
+
# hf_hub_download(repo_id="TheBloke/Mistral-7B-Instruct-v0.1-GGUF", local_dir=".", filename="mistral-7b-instruct-v0.1.Q5_K_M.gguf")
|
145 |
+
# mistral_model_path="./mistral-7b-instruct-v0.1.Q5_K_M.gguf"
|
146 |
+
|
147 |
+
#print("Downloading Yi-6B")
|
148 |
+
#Yi-6B
|
149 |
+
# Note current Yi is text-generation model not an instruct based model
|
150 |
+
#hf_hub_download(repo_id="TheBloke/Yi-6B-GGUF", local_dir=".", filename="yi-6b.Q5_K_M.gguf")
|
151 |
+
#yi_model_path="./yi-6b.Q5_K_M.gguf"
|
152 |
+
|
153 |
+
|
154 |
+
from llama_cpp import Llama
|
155 |
+
# set GPU_LAYERS to 15 if you have a 8GB GPU so both models can fit in
|
156 |
+
# else 35 full layers + XTTS works fine on T4 16GB
|
157 |
+
# 5gb per llm, 4gb XTTS -> full layers should fit T4 16GB , 2LLM + XTTS
|
158 |
+
GPU_LAYERS=int(os.environ.get("GPU_LAYERS",35))
|
159 |
+
|
160 |
+
LLM_STOP_WORDS= ["</s>","<|user|>","/s>","<EOT>","[/INST]"]
|
161 |
+
|
162 |
+
LLAMA_VERBOSE=False
|
163 |
+
# print("Running Mistral")
|
164 |
+
# llm_mistral = Llama(model_path=mistral_model_path,n_gpu_layers=GPU_LAYERS,max_new_tokens=256, context_window=4096, n_ctx=4096,n_batch=128,verbose=LLAMA_VERBOSE)
|
165 |
+
#print("Running LLM Mistral as InferenceClient")
|
166 |
+
#llm_mistral = InferenceClient("mistralai/Mistral-7B-Instruct-v0.1")
|
167 |
+
|
168 |
+
|
169 |
+
print("Running LLM Zephyr")
|
170 |
+
llm_zephyr = Llama(model_path=zephyr_model_path,n_gpu_layers=round(GPU_LAYERS/2),max_new_tokens=256, context_window=4096, n_ctx=4096,n_batch=128,verbose=LLAMA_VERBOSE)
|
171 |
+
|
172 |
+
#print("Running Yi LLM")
|
173 |
+
#llm_yi = Llama(model_path=yi_model_path,n_gpu_layers=round(GPU_LAYERS/2),max_new_tokens=256, context_window=4096, n_ctx=4096,n_batch=128,verbose=LLAMA_VERBOSE)
|
174 |
+
|
175 |
+
|
176 |
+
# # Mistral formatter
|
177 |
+
# def format_prompt_mistral(message, history, system_message=system_message,system_understand_message=system_understand_message):
|
178 |
+
# prompt = (
|
179 |
+
# "<s>[INST]" + system_message + "[/INST]" + system_understand_message + "</s>"
|
180 |
+
# )
|
181 |
+
# for user_prompt, bot_response in history:
|
182 |
+
# prompt += f"[INST] {user_prompt} [/INST]"
|
183 |
+
# prompt += f" {bot_response}</s> "
|
184 |
+
|
185 |
+
# if message=="":
|
186 |
+
# message="Hello"
|
187 |
+
# prompt += f"[INST] {message} [/INST]"
|
188 |
+
# return prompt
|
189 |
+
|
190 |
+
def format_prompt_yi(message, history, system_message=system_message,system_understand_message=system_understand_message):
|
191 |
+
prompt = (
|
192 |
+
"<s>[INST] [SYS]\n" + system_message + "\n[/SYS]\n\n[/INST]"
|
193 |
+
)
|
194 |
+
for user_prompt, bot_response in history:
|
195 |
+
prompt += f"[INST] {user_prompt} [/INST]"
|
196 |
+
prompt += f" {bot_response}</s> "
|
197 |
+
|
198 |
+
if message=="":
|
199 |
+
message="Hello"
|
200 |
+
prompt += f"[INST] {message} [/INST]"
|
201 |
+
return prompt
|
202 |
+
|
203 |
+
|
204 |
+
# <|system|>
|
205 |
+
# You are a friendly chatbot who always responds in the style of a pirate.</s>
|
206 |
+
# <|user|>
|
207 |
+
# How many helicopters can a human eat in one sitting?</s>
|
208 |
+
# <|assistant|>
|
209 |
+
# Ah, me hearty matey! But yer question be a puzzler! A human cannot eat a helicopter in one sitting, as helicopters are not edible. They be made of metal, plastic, and other materials, not food!
|
210 |
+
|
211 |
+
# Zephyr formatter
|
212 |
+
def format_prompt_zephyr(message, history, system_message=system_message):
|
213 |
+
prompt = (
|
214 |
+
"<|system|>\n" + system_message + "</s>"
|
215 |
+
)
|
216 |
+
for user_prompt, bot_response in history:
|
217 |
+
prompt += f"<|user|>\n{user_prompt}</s>"
|
218 |
+
prompt += f"<|assistant|>\n{bot_response}</s>"
|
219 |
+
if message=="":
|
220 |
+
message="Hello"
|
221 |
+
prompt += f"<|user|>\n{message}</s>"
|
222 |
+
prompt += f"<|assistant|>"
|
223 |
+
print(prompt)
|
224 |
+
return prompt
|
225 |
+
|
226 |
+
@spaces.GPU
|
227 |
+
def generate_local(
|
228 |
+
prompt,
|
229 |
+
history,
|
230 |
+
llm_model="zephyr",
|
231 |
+
system_message=None,
|
232 |
+
temperature=0.8,
|
233 |
+
max_tokens=256,
|
234 |
+
top_p=0.95,
|
235 |
+
stop = LLM_STOP_WORDS
|
236 |
+
):
|
237 |
+
temperature = float(temperature)
|
238 |
+
if temperature < 1e-2:
|
239 |
+
temperature = 1e-2
|
240 |
+
top_p = float(top_p)
|
241 |
+
|
242 |
+
generate_kwargs = dict(
|
243 |
+
temperature=temperature,
|
244 |
+
max_tokens=max_tokens,
|
245 |
+
top_p=top_p,
|
246 |
+
stop=stop
|
247 |
+
)
|
248 |
+
|
249 |
+
if "zephyr" in llm_model.lower():
|
250 |
+
sys_message= system_message.replace("##LLM_MODEL###","Zephyr").replace("##LLM_MODEL_PROVIDER###","Hugging Face")
|
251 |
+
formatted_prompt = format_prompt_zephyr(prompt, history,system_message=sys_message)
|
252 |
+
llm = llm_zephyr
|
253 |
+
else:
|
254 |
+
if "yi" in llm_model.lower():
|
255 |
+
llm_provider= "01.ai"
|
256 |
+
llm_model = "Yi"
|
257 |
+
llm = llm_yi
|
258 |
+
max_tokens= round(max_tokens/2)
|
259 |
+
else:
|
260 |
+
llm_provider= "Mistral"
|
261 |
+
llm_model = "Mistral"
|
262 |
+
llm = llm_mistral
|
263 |
+
sys_message= system_message.replace("##LLM_MODEL###",llm_model).replace("##LLM_MODEL_PROVIDER###",llm_provider)
|
264 |
+
sys_system_understand_message = system_understand_message.replace("##LLM_MODEL###",llm_model).replace("##LLM_MODEL_PROVIDER###",llm_provider)
|
265 |
+
|
266 |
+
if "yi" in llm_model.lower():
|
267 |
+
formatted_prompt = format_prompt_mistral(prompt, history,system_message=sys_message,system_understand_message="")
|
268 |
+
else:
|
269 |
+
formatted_prompt = format_prompt_mistral(prompt, history,system_message=sys_message,system_understand_message=sys_system_understand_message)
|
270 |
+
|
271 |
+
try:
|
272 |
+
print("LLM Input:", formatted_prompt)
|
273 |
+
if llm_model=="OTHER":
|
274 |
+
# Mistral endpoint too many Queues, wait time..
|
275 |
+
generate_kwargs = dict(
|
276 |
+
temperature=temperature,
|
277 |
+
max_new_tokens=max_tokens,
|
278 |
+
top_p=top_p,
|
279 |
+
)
|
280 |
+
|
281 |
+
stream = llm_mistral.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=False)
|
282 |
+
output = ""
|
283 |
+
for response in stream:
|
284 |
+
character = response.token.text
|
285 |
+
if character in LLM_STOP_WORDS:
|
286 |
+
# end of context
|
287 |
+
return
|
288 |
+
|
289 |
+
if emoji.is_emoji(character):
|
290 |
+
# Bad emoji not a meaning messes chat from next lines
|
291 |
+
return
|
292 |
+
|
293 |
+
output += character
|
294 |
+
yield output
|
295 |
+
else:
|
296 |
+
# Local GGUF
|
297 |
+
stream = llm(
|
298 |
+
formatted_prompt,
|
299 |
+
**generate_kwargs,
|
300 |
+
stream=True,
|
301 |
+
)
|
302 |
+
output = ""
|
303 |
+
for response in stream:
|
304 |
+
character= response["choices"][0]["text"]
|
305 |
+
|
306 |
+
if character in LLM_STOP_WORDS:
|
307 |
+
# end of context
|
308 |
+
return
|
309 |
+
|
310 |
+
if emoji.is_emoji(character):
|
311 |
+
# Bad emoji not a meaning messes chat from next lines
|
312 |
+
return
|
313 |
+
|
314 |
+
output += response["choices"][0]["text"].replace("<|assistant|>","").replace("<|user|>","")
|
315 |
+
yield output
|
316 |
+
|
317 |
+
except Exception as e:
|
318 |
+
if "Too Many Requests" in str(e):
|
319 |
+
print("ERROR: Too many requests on mistral client")
|
320 |
+
gr.Warning("Unfortunately Mistral is unable to process")
|
321 |
+
output = "Unfortuanately I am not able to process your request now !"
|
322 |
+
else:
|
323 |
+
print("Unhandled Exception: ", str(e))
|
324 |
+
gr.Warning("Unfortunately Mistral is unable to process")
|
325 |
+
output = "I do not know what happened but I could not understand you ."
|
326 |
+
|
327 |
+
return output
|
328 |
+
|
329 |
+
def get_latents(speaker_wav,voice_cleanup=False):
|
330 |
+
if (voice_cleanup):
|
331 |
+
try:
|
332 |
+
cleanup_filter="lowpass=8000,highpass=75,areverse,silenceremove=start_periods=1:start_silence=0:start_threshold=0.02,areverse,silenceremove=start_periods=1:start_silence=0:start_threshold=0.02"
|
333 |
+
resample_filter="-ac 1 -ar 22050"
|
334 |
+
out_filename = speaker_wav + str(uuid.uuid4()) + ".wav" #ffmpeg to know output format
|
335 |
+
#we will use newer ffmpeg as that has afftn denoise filter
|
336 |
+
shell_command = f"ffmpeg -y -i {speaker_wav} -af {cleanup_filter} {resample_filter} {out_filename}".split(" ")
|
337 |
+
|
338 |
+
command_result = subprocess.run([item for item in shell_command], capture_output=False,text=True, check=True)
|
339 |
+
speaker_wav=out_filename
|
340 |
+
print("Filtered microphone input")
|
341 |
+
except subprocess.CalledProcessError:
|
342 |
+
# There was an error - command exited with non-zero code
|
343 |
+
print("Error: failed filtering, use original microphone input")
|
344 |
+
else:
|
345 |
+
speaker_wav=speaker_wav
|
346 |
+
|
347 |
+
# create as function as we can populate here with voice cleanup/filtering
|
348 |
+
(
|
349 |
+
gpt_cond_latent,
|
350 |
+
speaker_embedding,
|
351 |
+
) = model.get_conditioning_latents(audio_path=speaker_wav)
|
352 |
+
return gpt_cond_latent, speaker_embedding
|
353 |
+
|
354 |
+
def wave_header_chunk(frame_input=b"", channels=1, sample_width=2, sample_rate=24000):
|
355 |
+
# This will create a wave header then append the frame input
|
356 |
+
# It should be first on a streaming wav file
|
357 |
+
# Other frames better should not have it (else you will hear some artifacts each chunk start)
|
358 |
+
wav_buf = io.BytesIO()
|
359 |
+
with wave.open(wav_buf, "wb") as vfout:
|
360 |
+
vfout.setnchannels(channels)
|
361 |
+
vfout.setsampwidth(sample_width)
|
362 |
+
vfout.setframerate(sample_rate)
|
363 |
+
vfout.writeframes(frame_input)
|
364 |
+
|
365 |
+
wav_buf.seek(0)
|
366 |
+
return wav_buf.read()
|
367 |
+
|
368 |
+
|
369 |
+
#Config will have more correct languages, they may be added before we append here
|
370 |
+
##["en","es","fr","de","it","pt","pl","tr","ru","nl","cs","ar","zh-cn","ja"]
|
371 |
+
|
372 |
+
xtts_supported_languages=config.languages
|
373 |
+
# def detect_language(prompt):
|
374 |
+
# # Fast language autodetection
|
375 |
+
# if len(prompt)>15:
|
376 |
+
# language_predicted=langid.classify(prompt)[0].strip() # strip need as there is space at end!
|
377 |
+
# if language_predicted == "zh":
|
378 |
+
# #we use zh-cn on xtts
|
379 |
+
# language_predicted = "zh-cn"
|
380 |
+
|
381 |
+
# if language_predicted not in xtts_supported_languages:
|
382 |
+
# print(f"Detected a language not supported by xtts :{language_predicted}, switching to english for now")
|
383 |
+
# gr.Warning(f"Language detected '{language_predicted}' can not be spoken properly 'yet' ")
|
384 |
+
# language= "en"
|
385 |
+
# else:
|
386 |
+
# language = language_predicted
|
387 |
+
# print(f"Language: Predicted sentence language:{language_predicted} , using language for xtts:{language}")
|
388 |
+
# else:
|
389 |
+
# # Hard to detect language fast in short sentence, use english default
|
390 |
+
# language = "en"
|
391 |
+
# print(f"Language: Prompt is short or autodetect language disabled using english for xtts")
|
392 |
+
|
393 |
+
# return language
|
394 |
+
|
395 |
+
def get_voice_streaming(prompt, language, latent_tuple, suffix="0"):
|
396 |
+
gpt_cond_latent, speaker_embedding = latent_tuple
|
397 |
+
|
398 |
+
try:
|
399 |
+
t0 = time.time()
|
400 |
+
chunks = model.inference_stream(
|
401 |
+
prompt,
|
402 |
+
language,
|
403 |
+
gpt_cond_latent,
|
404 |
+
speaker_embedding,
|
405 |
+
repetition_penalty=7.0,
|
406 |
+
temperature=0.85,
|
407 |
+
)
|
408 |
+
|
409 |
+
first_chunk = True
|
410 |
+
for i, chunk in enumerate(chunks):
|
411 |
+
if first_chunk:
|
412 |
+
first_chunk_time = time.time() - t0
|
413 |
+
metrics_text = f"Latency to first audio chunk: {round(first_chunk_time*1000)} milliseconds\n"
|
414 |
+
first_chunk = False
|
415 |
+
#print(f"Received chunk {i} of audio length {chunk.shape[-1]}")
|
416 |
+
|
417 |
+
# In case output is required to be multiple voice files
|
418 |
+
# out_file = f'{char}_{i}.wav'
|
419 |
+
# write(out_file, 24000, chunk.detach().cpu().numpy().squeeze())
|
420 |
+
# audio = AudioSegment.from_file(out_file)
|
421 |
+
# audio.export(out_file, format='wav')
|
422 |
+
# return out_file
|
423 |
+
# directly return chunk as bytes for streaming
|
424 |
+
chunk = chunk.detach().cpu().numpy().squeeze()
|
425 |
+
chunk = (chunk * 32767).astype(np.int16)
|
426 |
+
|
427 |
+
yield chunk.tobytes()
|
428 |
+
|
429 |
+
except RuntimeError as e:
|
430 |
+
if "device-side assert" in str(e):
|
431 |
+
# cannot do anything on cuda device side error, need tor estart
|
432 |
+
print(
|
433 |
+
f"Exit due to: Unrecoverable exception caused by prompt:{prompt}",
|
434 |
+
flush=True,
|
435 |
+
)
|
436 |
+
gr.Warning("Unhandled Exception encounter, please retry in a minute")
|
437 |
+
print("Cuda device-assert Runtime encountered need restart")
|
438 |
+
|
439 |
+
# HF Space specific.. This error is unrecoverable need to restart space
|
440 |
+
api.restart_space(repo_id=repo_id)
|
441 |
+
else:
|
442 |
+
print("RuntimeError: non device-side assert error:", str(e))
|
443 |
+
# Does not require warning happens on empty chunk and at end
|
444 |
+
###gr.Warning("Unhandled Exception encounter, please retry in a minute")
|
445 |
+
return None
|
446 |
+
return None
|
447 |
+
except:
|
448 |
+
return None
|
449 |
+
|
450 |
+
|
451 |
+
def transcribe(wav_path):
|
452 |
+
try:
|
453 |
+
# get result from whisper and strip it to delete begin and end space
|
454 |
+
return whisper_client.predict(
|
455 |
+
wav_path, # str (filepath or URL to file) in 'inputs' Audio component
|
456 |
+
"transcribe", # str in 'Task' Radio component
|
457 |
+
api_name="/predict"
|
458 |
+
).strip()
|
459 |
+
except:
|
460 |
+
gr.Warning("There was a problem with Whisper endpoint, telling a joke for you.")
|
461 |
+
return "There was a problem with my voice, tell me joke"
|
462 |
+
|
463 |
+
|
464 |
+
# Will be triggered on text submit (will send to generate_speech)
|
465 |
+
def add_text(history, text):
|
466 |
+
history = [] if history is None else history
|
467 |
+
history = history + [(text, None)]
|
468 |
+
return history, gr.update(value="", interactive=False)
|
469 |
+
|
470 |
+
# Will be triggered on voice submit (will transribe and send to generate_speech)
|
471 |
+
def add_file(history, file):
|
472 |
+
history = [] if history is None else history
|
473 |
+
|
474 |
+
try:
|
475 |
+
text = transcribe(file)
|
476 |
+
print("Transcribed text:", text)
|
477 |
+
except Exception as e:
|
478 |
+
print(str(e))
|
479 |
+
gr.Warning("There was an issue with transcription, please try writing for now")
|
480 |
+
# Apply a null text on error
|
481 |
+
text = "Transcription seems failed, please tell me a joke about chickens"
|
482 |
+
|
483 |
+
history = history + [(text, None)]
|
484 |
+
return history, gr.update(value="", interactive=False)
|
485 |
+
|
486 |
+
|
487 |
+
##NOTE: not using this as it yields a chacter each time while we need to feed history to TTS
|
488 |
+
def bot(history, system_prompt=""):
|
489 |
+
history = [["", None]] if history is None else history
|
490 |
+
|
491 |
+
if system_prompt == "":
|
492 |
+
system_prompt = system_message
|
493 |
+
|
494 |
+
history[-1][1] = ""
|
495 |
+
for character in generate(history[-1][0], history[:-1]):
|
496 |
+
history[-1][1] = character
|
497 |
+
yield history
|
498 |
+
|
499 |
+
|
500 |
+
def get_sentence(history, chatbot_role,llm_model,system_prompt=""):
|
501 |
+
|
502 |
+
history = [["", None]] if history is None else history
|
503 |
+
|
504 |
+
if system_prompt == "":
|
505 |
+
system_prompt = system_message
|
506 |
+
|
507 |
+
history[-1][1] = ""
|
508 |
+
|
509 |
+
mistral_start = time.time()
|
510 |
+
|
511 |
+
sentence_list = []
|
512 |
+
sentence_hash_list = []
|
513 |
+
|
514 |
+
text_to_generate = ""
|
515 |
+
stored_sentence = None
|
516 |
+
stored_sentence_hash = None
|
517 |
+
|
518 |
+
print(chatbot_role)
|
519 |
+
print(llm_model)
|
520 |
+
|
521 |
+
for character in generate_local(history[-1][0], history[:-1],system_message=ROLE_PROMPTS[chatbot_role],llm_model=llm_model):
|
522 |
+
history[-1][1] = character.replace("<|assistant|>","")
|
523 |
+
# It is coming word by word
|
524 |
+
|
525 |
+
text_to_generate = nltk.sent_tokenize(history[-1][1].replace("\n", " ").replace("<|assistant|>"," ").replace("<|ass>","").replace("[/ASST]","").replace("[/ASSI]","").replace("[/ASS]","").replace("","").strip())
|
526 |
+
if len(text_to_generate) > 1:
|
527 |
+
|
528 |
+
dif = len(text_to_generate) - len(sentence_list)
|
529 |
+
|
530 |
+
if dif == 1 and len(sentence_list) != 0:
|
531 |
+
continue
|
532 |
+
|
533 |
+
if dif == 2 and len(sentence_list) != 0 and stored_sentence is not None:
|
534 |
+
continue
|
535 |
+
|
536 |
+
# All this complexity due to trying append first short sentence to next one for proper language auto-detect
|
537 |
+
if stored_sentence is not None and stored_sentence_hash is None and dif>1:
|
538 |
+
#means we consumed stored sentence and should look at next sentence to generate
|
539 |
+
sentence = text_to_generate[len(sentence_list)+1]
|
540 |
+
elif stored_sentence is not None and len(text_to_generate)>2 and stored_sentence_hash is not None:
|
541 |
+
print("Appending stored")
|
542 |
+
sentence = stored_sentence + text_to_generate[len(sentence_list)+1]
|
543 |
+
stored_sentence_hash = None
|
544 |
+
else:
|
545 |
+
sentence = text_to_generate[len(sentence_list)]
|
546 |
+
|
547 |
+
# too short sentence just append to next one if there is any
|
548 |
+
# this is for proper language detection
|
549 |
+
if len(sentence)<=15 and stored_sentence_hash is None and stored_sentence is None:
|
550 |
+
if sentence[-1] in [".","!","?"]:
|
551 |
+
if stored_sentence_hash != hash(sentence):
|
552 |
+
stored_sentence = sentence
|
553 |
+
stored_sentence_hash = hash(sentence)
|
554 |
+
print("Storing:",stored_sentence)
|
555 |
+
continue
|
556 |
+
|
557 |
+
|
558 |
+
sentence_hash = hash(sentence)
|
559 |
+
if stored_sentence_hash is not None and sentence_hash == stored_sentence_hash:
|
560 |
+
continue
|
561 |
+
|
562 |
+
if sentence_hash not in sentence_hash_list:
|
563 |
+
sentence_hash_list.append(sentence_hash)
|
564 |
+
sentence_list.append(sentence)
|
565 |
+
print("New Sentence: ", sentence)
|
566 |
+
yield (sentence, history)
|
567 |
+
|
568 |
+
# return that final sentence token
|
569 |
+
try:
|
570 |
+
last_sentence = nltk.sent_tokenize(history[-1][1].replace("\n", " ").replace("<|ass>","").replace("[/ASST]","").replace("[/ASSI]","").replace("[/ASS]","").replace("","").strip())[-1]
|
571 |
+
sentence_hash = hash(last_sentence)
|
572 |
+
if sentence_hash not in sentence_hash_list:
|
573 |
+
if stored_sentence is not None and stored_sentence_hash is not None:
|
574 |
+
last_sentence = stored_sentence + last_sentence
|
575 |
+
stored_sentence = stored_sentence_hash = None
|
576 |
+
print("Last Sentence with stored:",last_sentence)
|
577 |
+
|
578 |
+
sentence_hash_list.append(sentence_hash)
|
579 |
+
sentence_list.append(last_sentence)
|
580 |
+
print("Last Sentence: ", last_sentence)
|
581 |
+
|
582 |
+
yield (last_sentence, history)
|
583 |
+
except:
|
584 |
+
print("ERROR on last sentence history is :", history)
|
585 |
+
|
586 |
+
|
587 |
+
from scipy.io.wavfile import write
|
588 |
+
from pydub import AudioSegment
|
589 |
+
|
590 |
+
second_of_silence = AudioSegment.silent() # use default
|
591 |
+
second_of_silence.export("sil.wav", format='wav')
|
592 |
+
|
593 |
+
|
594 |
+
def generate_speech(history,chatbot_role,llm_model):
|
595 |
+
# Must set autoplay to True first
|
596 |
+
yield (history, chatbot_role, "", wave_header_chunk() )
|
597 |
+
for sentence, history in get_sentence(history,chatbot_role,llm_model):
|
598 |
+
if sentence != "":
|
599 |
+
print("BG: inserting sentence to queue")
|
600 |
+
|
601 |
+
generated_speech = generate_speech_for_sentence(history, chatbot_role, sentence,return_as_byte=True)
|
602 |
+
if generated_speech is not None:
|
603 |
+
_, audio_dict = generated_speech
|
604 |
+
# We are using byte streaming
|
605 |
+
yield (history, chatbot_role, sentence, audio_dict["value"] )
|
606 |
+
|
607 |
+
|
608 |
+
# will generate speech audio file per sentence
|
609 |
+
def generate_speech_for_sentence(history, chatbot_role, sentence, return_as_byte=False):
|
610 |
+
language = "autodetect"
|
611 |
+
|
612 |
+
wav_bytestream = b""
|
613 |
+
|
614 |
+
if len(sentence)==0:
|
615 |
+
print("EMPTY SENTENCE")
|
616 |
+
return
|
617 |
+
|
618 |
+
# Sometimes prompt </s> coming on output remove it
|
619 |
+
# Some post process for speech only
|
620 |
+
sentence = sentence.replace("</s>", "")
|
621 |
+
# remove code from speech
|
622 |
+
sentence = re.sub("```.*```", "", sentence, flags=re.DOTALL)
|
623 |
+
sentence = re.sub("`.*`", "", sentence, flags=re.DOTALL)
|
624 |
+
|
625 |
+
sentence = re.sub("\(.*\)", "", sentence, flags=re.DOTALL)
|
626 |
+
|
627 |
+
sentence = sentence.replace("```", "")
|
628 |
+
sentence = sentence.replace("...", " ")
|
629 |
+
sentence = sentence.replace("(", " ")
|
630 |
+
sentence = sentence.replace(")", " ")
|
631 |
+
sentence = sentence.replace("<|assistant|>","")
|
632 |
+
|
633 |
+
if len(sentence)==0:
|
634 |
+
print("EMPTY SENTENCE after processing")
|
635 |
+
return
|
636 |
+
|
637 |
+
# A fast fix for last chacter, may produce weird sounds if it is with text
|
638 |
+
#if (sentence[-1] in ["!", "?", ".", ","]) or (sentence[-2] in ["!", "?", ".", ","]):
|
639 |
+
# # just add a space
|
640 |
+
# sentence = sentence[:-1] + " " + sentence[-1]
|
641 |
+
|
642 |
+
# regex does the job well
|
643 |
+
sentence= re.sub("([^\x00-\x7F]|\w)(\.|\γ|\?|\!)",r"\1 \2\2",sentence)
|
644 |
+
|
645 |
+
print("Sentence for speech:", sentence)
|
646 |
+
|
647 |
+
|
648 |
+
try:
|
649 |
+
SENTENCE_SPLIT_LENGTH=350
|
650 |
+
if len(sentence)<SENTENCE_SPLIT_LENGTH:
|
651 |
+
# no problem continue on
|
652 |
+
sentence_list = [sentence]
|
653 |
+
else:
|
654 |
+
# Until now nltk likely split sentences properly but we need additional
|
655 |
+
# check for longer sentence and split at last possible position
|
656 |
+
# Do whatever necessary, first break at hypens then spaces and then even split very long words
|
657 |
+
sentence_list=textwrap.wrap(sentence,SENTENCE_SPLIT_LENGTH)
|
658 |
+
print("SPLITTED LONG SENTENCE:",sentence_list)
|
659 |
+
|
660 |
+
for sentence in sentence_list:
|
661 |
+
|
662 |
+
if any(c.isalnum() for c in sentence):
|
663 |
+
if language=="autodetect":
|
664 |
+
#on first call autodetect, nexts sentence calls will use same language
|
665 |
+
language = language = "en" #detect_language(sentence)
|
666 |
+
|
667 |
+
#exists at least 1 alphanumeric (utf-8)
|
668 |
+
audio_stream = get_voice_streaming(
|
669 |
+
sentence, language, latent_map[chatbot_role]
|
670 |
+
)
|
671 |
+
else:
|
672 |
+
# likely got a ' or " or some other text without alphanumeric in it
|
673 |
+
audio_stream = None
|
674 |
+
|
675 |
+
# XTTS is actually using streaming response but we are playing audio by sentence
|
676 |
+
# If you want direct XTTS voice streaming (send each chunk to voice ) you may set DIRECT_STREAM=1 environment variable
|
677 |
+
if audio_stream is not None:
|
678 |
+
frame_length = 0
|
679 |
+
for chunk in audio_stream:
|
680 |
+
try:
|
681 |
+
wav_bytestream += chunk
|
682 |
+
frame_length += len(chunk)
|
683 |
+
except:
|
684 |
+
# hack to continue on playing. sometimes last chunk is empty , will be fixed on next TTS
|
685 |
+
continue
|
686 |
+
|
687 |
+
# Filter output for better voice
|
688 |
+
filter_output=False
|
689 |
+
if filter_output:
|
690 |
+
data_s16 = np.frombuffer(wav_bytestream, dtype=np.int16, count=len(wav_bytestream)//2, offset=0)
|
691 |
+
float_data = data_s16 * 0.5**15
|
692 |
+
reduced_noise = nr.reduce_noise(y=float_data, sr=24000,prop_decrease =0.8,n_fft=1024)
|
693 |
+
wav_bytestream = (reduced_noise * 32767).astype(np.int16)
|
694 |
+
wav_bytestream = wav_bytestream.tobytes()
|
695 |
+
|
696 |
+
if audio_stream is not None:
|
697 |
+
if not return_as_byte:
|
698 |
+
audio_unique_filename = "/tmp/"+ str(uuid.uuid4())+".wav"
|
699 |
+
with wave.open(audio_unique_filename, "w") as f:
|
700 |
+
f.setnchannels(1)
|
701 |
+
# 2 bytes per sample.
|
702 |
+
f.setsampwidth(2)
|
703 |
+
f.setframerate(24000)
|
704 |
+
f.writeframes(wav_bytestream)
|
705 |
+
|
706 |
+
return (history , gr.Audio.update(value=audio_unique_filename, autoplay=True))
|
707 |
+
else:
|
708 |
+
return (history , gr.Audio.update(value=wav_bytestream, autoplay=True))
|
709 |
+
except RuntimeError as e:
|
710 |
+
if "device-side assert" in str(e):
|
711 |
+
# cannot do anything on cuda device side error, need tor estart
|
712 |
+
print(
|
713 |
+
f"Exit due to: Unrecoverable exception caused by prompt:{sentence}",
|
714 |
+
flush=True,
|
715 |
+
)
|
716 |
+
gr.Warning("Unhandled Exception encounter, please retry in a minute")
|
717 |
+
print("Cuda device-assert Runtime encountered need restart")
|
718 |
+
|
719 |
+
# HF Space specific.. This error is unrecoverable need to restart space
|
720 |
+
api.restart_space(repo_id=repo_id)
|
721 |
+
else:
|
722 |
+
print("RuntimeError: non device-side assert error:", str(e))
|
723 |
+
raise e
|
724 |
+
|
725 |
+
print("All speech ended")
|
726 |
+
return
|
727 |
+
|
728 |
+
latent_map = {}
|
729 |
+
latent_map["AI Assistant"] = get_latents("examples/female.wav")
|
730 |
+
latent_map["AI Beard The Pirate"] = get_latents("examples/pirate_by_coqui.wav")
|
731 |
+
|
732 |
+
#### GRADIO INTERFACE ####
|
733 |
+
|
734 |
+
EXAMPLES = [
|
735 |
+
[[],"AI Assistant","Write a 100-word article on 'Benefits of Open-Source in AI research"],
|
736 |
+
]
|
737 |
+
|
738 |
+
#MODELS = ["Zephyr 7B Beta","Mistral 7B Instruct"]
|
739 |
+
MODELS = ["Zephyr 7B Beta"]
|
740 |
+
OTHER_HTML=f""""""
|
741 |
+
|
742 |
+
with gr.Blocks(title=title) as demo:
|
743 |
+
|
744 |
+
with gr.Row():
|
745 |
+
model_selected = gr.Dropdown(
|
746 |
+
label="LLM Model",
|
747 |
+
#info="Mistral, Zephyr: Mistral uses inference endpoint, Zephyr is 5 bit GGUF",
|
748 |
+
choices=MODELS,
|
749 |
+
max_choices=1,
|
750 |
+
value=MODELS[0],
|
751 |
+
visible = False,
|
752 |
+
)
|
753 |
+
chatbot = gr.Chatbot(
|
754 |
+
[],
|
755 |
+
elem_id="chatbot",
|
756 |
+
#avatar_images=("examples/hf-logo.png", "examples/coqui-logo.png"),
|
757 |
+
bubble_full_width=False,
|
758 |
+
)
|
759 |
+
with gr.Row():
|
760 |
+
chatbot_role = gr.Dropdown(
|
761 |
+
label="Role of the Chatbot",
|
762 |
+
info="How should Chatbot talk like",
|
763 |
+
choices=ROLES,
|
764 |
+
max_choices=1,
|
765 |
+
value=ROLES[0],
|
766 |
+
visible = False,
|
767 |
+
)
|
768 |
+
with gr.Row():
|
769 |
+
txt = gr.Textbox(
|
770 |
+
scale=3,
|
771 |
+
show_label=False,
|
772 |
+
placeholder="Enter text and press enter, or speak to your microphone",
|
773 |
+
container=False,
|
774 |
+
interactive=True,
|
775 |
+
)
|
776 |
+
txt_btn = gr.Button(value="Submit text", scale=1)
|
777 |
+
btn = gr.Audio(source="microphone", type="filepath", scale=4)
|
778 |
+
|
779 |
+
def stop():
|
780 |
+
print("Audio STOP")
|
781 |
+
set_audio_playing(False)
|
782 |
+
|
783 |
+
with gr.Row():
|
784 |
+
sentence = gr.Textbox(visible=False)
|
785 |
+
audio = gr.Audio(
|
786 |
+
value=None,
|
787 |
+
label="Generated audio response",
|
788 |
+
streaming=True,
|
789 |
+
autoplay=True,
|
790 |
+
interactive=False,
|
791 |
+
show_label=True,
|
792 |
+
visible = False,
|
793 |
+
)
|
794 |
+
|
795 |
+
audio.end(stop)
|
796 |
+
|
797 |
+
with gr.Row():
|
798 |
+
gr.Examples(
|
799 |
+
EXAMPLES,
|
800 |
+
[chatbot,chatbot_role, txt],
|
801 |
+
[chatbot,chatbot_role, txt],
|
802 |
+
add_text,
|
803 |
+
cache_examples=False,
|
804 |
+
run_on_click=False, # Will not work , user should submit it
|
805 |
+
)
|
806 |
+
|
807 |
+
def clear_inputs(chatbot):
|
808 |
+
return None
|
809 |
+
clear_btn = gr.ClearButton([chatbot, audio])
|
810 |
+
chatbot_role.change(fn=clear_inputs, inputs=[chatbot], outputs=[chatbot])
|
811 |
+
model_selected.change(fn=clear_inputs, inputs=[chatbot], outputs=[chatbot])
|
812 |
+
|
813 |
+
txt_msg = txt_btn.click(add_text, [chatbot, txt], [chatbot, txt], queue=False).then(
|
814 |
+
generate_speech, [chatbot,chatbot_role,model_selected], [chatbot,chatbot_role, sentence, audio]
|
815 |
+
)
|
816 |
+
|
817 |
+
txt_msg.then(lambda: gr.update(interactive=True), None, [txt], queue=False)
|
818 |
+
|
819 |
+
txt_msg = txt.submit(add_text, [chatbot, txt], [chatbot, txt], queue=False).then(
|
820 |
+
generate_speech, [chatbot,chatbot_role,model_selected], [chatbot,chatbot_role, sentence, audio]
|
821 |
+
)
|
822 |
+
|
823 |
+
txt_msg.then(lambda: gr.update(interactive=True), None, [txt], queue=False)
|
824 |
+
|
825 |
+
file_msg = btn.stop_recording(
|
826 |
+
add_file, [chatbot, btn], [chatbot, txt], queue=False
|
827 |
+
).then(
|
828 |
+
generate_speech, [chatbot,chatbot_role,model_selected], [chatbot,chatbot_role, sentence, audio]
|
829 |
+
)
|
830 |
+
|
831 |
+
file_msg.then(lambda: (gr.update(interactive=True),gr.update(interactive=True,value=None)), None, [txt, btn], queue=False)
|
832 |
+
|
833 |
+
|
834 |
+
demo.queue()
|
835 |
+
demo.launch(debug=True,share=True)
|