xtts-and-whisper-update (#3)
Browse files- xtts and whisper jax (145f28e71d0018b8125e2a088a6e2fc7efff9ce9)
- remove unneded testfile (4d98613abf7d496dd3b12d2af377f8ba4ed587f4)
- add ffmpeg import (0e056f7dc33bdc6d3eab93d5ded7bb4ef630e102)
- app.py +133 -31
- requirements.txt +5 -2
app.py
CHANGED
@@ -11,8 +11,36 @@ import nltk # we'll use this to split into sentences
|
|
11 |
nltk.download('punkt')
|
12 |
import uuid
|
13 |
|
|
|
|
|
|
|
14 |
from TTS.api import TTS
|
15 |
-
tts
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
|
17 |
title = "Voice chat with Mistral 7B Instruct"
|
18 |
|
@@ -44,11 +72,20 @@ from gradio_client import Client
|
|
44 |
from huggingface_hub import InferenceClient
|
45 |
|
46 |
|
47 |
-
|
|
|
|
|
|
|
48 |
text_client = InferenceClient(
|
49 |
"mistralai/Mistral-7B-Instruct-v0.1"
|
50 |
)
|
51 |
|
|
|
|
|
|
|
|
|
|
|
|
|
52 |
|
53 |
def format_prompt(message, history):
|
54 |
prompt = "<s>"
|
@@ -77,22 +114,35 @@ def generate(
|
|
77 |
|
78 |
formatted_prompt = format_prompt(prompt, history)
|
79 |
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
86 |
return output
|
87 |
|
88 |
|
89 |
def transcribe(wav_path):
|
90 |
|
|
|
91 |
return whisper_client.predict(
|
92 |
wav_path, # str (filepath or URL to file) in 'inputs' Audio component
|
93 |
"transcribe", # str in 'Task' Radio component
|
|
|
94 |
api_name="/predict"
|
95 |
-
)
|
96 |
|
97 |
|
98 |
# Chatbot demo with multimodal input (text, markdown, LaTeX, code blocks, image, audio, & video). Plus shows support for streaming text.
|
@@ -106,9 +156,17 @@ def add_text(history, text):
|
|
106 |
|
107 |
def add_file(history, file):
|
108 |
history = [] if history is None else history
|
109 |
-
|
110 |
-
|
111 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
112 |
|
113 |
history = history + [(text, None)]
|
114 |
return history
|
@@ -126,29 +184,65 @@ def bot(history, system_prompt=""):
|
|
126 |
history[-1][1] = character
|
127 |
yield history
|
128 |
|
129 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
130 |
def generate_speech(history):
|
131 |
text_to_generate = history[-1][1]
|
132 |
text_to_generate = text_to_generate.replace("\n", " ").strip()
|
133 |
text_to_generate = nltk.sent_tokenize(text_to_generate)
|
134 |
-
|
135 |
-
filename = f"{uuid.uuid4()}.wav"
|
136 |
-
sampling_rate = tts.synthesizer.tts_config.audio["sample_rate"]
|
137 |
-
silence = [0] * int(0.25 * sampling_rate)
|
138 |
|
139 |
-
|
140 |
-
for sentence in text_to_generate:
|
141 |
-
try:
|
142 |
|
143 |
-
|
144 |
-
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
150 |
|
151 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
152 |
|
153 |
except RuntimeError as e :
|
154 |
if "device-side assert" in str(e):
|
@@ -163,6 +257,14 @@ def generate_speech(history):
|
|
163 |
else:
|
164 |
print("RuntimeError: non device-side assert error:", str(e))
|
165 |
raise e
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
166 |
|
167 |
with gr.Blocks(title=title) as demo:
|
168 |
gr.Markdown(DESCRIPTION)
|
@@ -186,7 +288,7 @@ with gr.Blocks(title=title) as demo:
|
|
186 |
btn = gr.Audio(source="microphone", type="filepath", scale=4)
|
187 |
|
188 |
with gr.Row():
|
189 |
-
audio = gr.Audio(type="numpy", streaming=
|
190 |
|
191 |
clear_btn = gr.ClearButton([chatbot, audio])
|
192 |
|
@@ -210,11 +312,11 @@ with gr.Blocks(title=title) as demo:
|
|
210 |
gr.Markdown("""
|
211 |
This Space demonstrates how to speak to a chatbot, based solely on open-source models.
|
212 |
It relies on 3 models:
|
213 |
-
1. [Whisper-large-v2](https://huggingface.co/spaces/sanchit-gandhi/whisper-
|
214 |
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).
|
215 |
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.
|
216 |
|
217 |
Note:
|
218 |
- By using this demo you agree to the terms of the Coqui Public Model License at https://coqui.ai/cpml""")
|
219 |
demo.queue()
|
220 |
-
demo.launch(debug=True)
|
|
|
11 |
nltk.download('punkt')
|
12 |
import uuid
|
13 |
|
14 |
+
import ffmpeg
|
15 |
+
import librosa
|
16 |
+
import torchaudio
|
17 |
from TTS.api import TTS
|
18 |
+
from TTS.tts.configs.xtts_config import XttsConfig
|
19 |
+
from TTS.tts.models.xtts import Xtts
|
20 |
+
from TTS.utils.generic_utils import get_user_data_dir
|
21 |
+
|
22 |
+
# This will trigger downloading model
|
23 |
+
print("Downloading if not downloaded Coqui XTTS V1")
|
24 |
+
tts = TTS("tts_models/multilingual/multi-dataset/xtts_v1")
|
25 |
+
del tts
|
26 |
+
print("XTTS downloaded")
|
27 |
+
|
28 |
+
print("Loading XTTS")
|
29 |
+
#Below will use model directly for inference
|
30 |
+
model_path = os.path.join(get_user_data_dir("tts"), "tts_models--multilingual--multi-dataset--xtts_v1")
|
31 |
+
config = XttsConfig()
|
32 |
+
config.load_json(os.path.join(model_path, "config.json"))
|
33 |
+
model = Xtts.init_from_config(config)
|
34 |
+
model.load_checkpoint(
|
35 |
+
config,
|
36 |
+
checkpoint_path=os.path.join(model_path, "model.pth"),
|
37 |
+
vocab_path=os.path.join(model_path, "vocab.json"),
|
38 |
+
eval=True,
|
39 |
+
use_deepspeed=True
|
40 |
+
)
|
41 |
+
model.cuda()
|
42 |
+
print("Done loading TTS")
|
43 |
+
|
44 |
|
45 |
title = "Voice chat with Mistral 7B Instruct"
|
46 |
|
|
|
72 |
from huggingface_hub import InferenceClient
|
73 |
|
74 |
|
75 |
+
# This client is down
|
76 |
+
#whisper_client = Client("https://sanchit-gandhi-whisper-large-v2.hf.space/")
|
77 |
+
# Replacement whisper client, it may be time limited
|
78 |
+
whisper_client = Client("https://sanchit-gandhi-whisper-jax.hf.space")
|
79 |
text_client = InferenceClient(
|
80 |
"mistralai/Mistral-7B-Instruct-v0.1"
|
81 |
)
|
82 |
|
83 |
+
###### COQUI TTS FUNCTIONS ######
|
84 |
+
def get_latents(speaker_wav):
|
85 |
+
# create as function as we can populate here with voice cleanup/filtering
|
86 |
+
gpt_cond_latent, diffusion_conditioning, speaker_embedding = model.get_conditioning_latents(audio_path=speaker_wav)
|
87 |
+
return gpt_cond_latent, diffusion_conditioning, speaker_embedding
|
88 |
+
|
89 |
|
90 |
def format_prompt(message, history):
|
91 |
prompt = "<s>"
|
|
|
114 |
|
115 |
formatted_prompt = format_prompt(prompt, history)
|
116 |
|
117 |
+
try:
|
118 |
+
stream = text_client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=False)
|
119 |
+
output = ""
|
120 |
+
for response in stream:
|
121 |
+
output += response.token.text
|
122 |
+
yield output
|
123 |
+
|
124 |
+
except Exception as e:
|
125 |
+
if "Too Many Requests" in str(e):
|
126 |
+
print("ERROR: Too many requests on mistral client")
|
127 |
+
gr.Warning("Unfortunately Mistral is unable to process")
|
128 |
+
output = "Unfortuanately I am not able to process your request now !"
|
129 |
+
else:
|
130 |
+
print("Unhandled Exception: ", str(e))
|
131 |
+
gr.Warning("Unfortunately Mistral is unable to process")
|
132 |
+
output = "I do not know what happened but I could not understand you ."
|
133 |
+
|
134 |
return output
|
135 |
|
136 |
|
137 |
def transcribe(wav_path):
|
138 |
|
139 |
+
# get first element from whisper_jax and strip it to delete begin and end space
|
140 |
return whisper_client.predict(
|
141 |
wav_path, # str (filepath or URL to file) in 'inputs' Audio component
|
142 |
"transcribe", # str in 'Task' Radio component
|
143 |
+
False, # return_timestamps=False for whisper-jax https://gist.github.com/sanchit-gandhi/781dd7003c5b201bfe16d28634c8d4cf#file-whisper_jax_endpoint-py
|
144 |
api_name="/predict"
|
145 |
+
)[0].strip()
|
146 |
|
147 |
|
148 |
# Chatbot demo with multimodal input (text, markdown, LaTeX, code blocks, image, audio, & video). Plus shows support for streaming text.
|
|
|
156 |
|
157 |
def add_file(history, file):
|
158 |
history = [] if history is None else history
|
159 |
+
|
160 |
+
try:
|
161 |
+
text = transcribe(
|
162 |
+
file
|
163 |
+
)
|
164 |
+
print("Transcribed text:",text)
|
165 |
+
except Exception as e:
|
166 |
+
print(str(e))
|
167 |
+
gr.Warning("There was an issue with transcription, please try writing for now")
|
168 |
+
# Apply a null text on error
|
169 |
+
text = "Transcription seems failed, please tell me a joke about chickens"
|
170 |
|
171 |
history = history + [(text, None)]
|
172 |
return history
|
|
|
184 |
history[-1][1] = character
|
185 |
yield history
|
186 |
|
187 |
+
|
188 |
+
def get_latents(speaker_wav):
|
189 |
+
# Generate speaker embedding and latents for TTS
|
190 |
+
gpt_cond_latent, diffusion_conditioning, speaker_embedding = model.get_conditioning_latents(audio_path=speaker_wav)
|
191 |
+
return gpt_cond_latent, diffusion_conditioning, speaker_embedding
|
192 |
+
|
193 |
+
latent_map={}
|
194 |
+
latent_map["Female_Voice"] = get_latents("examples/female.wav")
|
195 |
+
|
196 |
+
def get_voice(prompt,language, latent_tuple,suffix="0"):
|
197 |
+
gpt_cond_latent,diffusion_conditioning, speaker_embedding = latent_tuple
|
198 |
+
# Direct version
|
199 |
+
t0 = time.time()
|
200 |
+
out = model.inference(
|
201 |
+
prompt,
|
202 |
+
language,
|
203 |
+
gpt_cond_latent,
|
204 |
+
speaker_embedding,
|
205 |
+
diffusion_conditioning
|
206 |
+
)
|
207 |
+
inference_time = time.time() - t0
|
208 |
+
print(f"I: Time to generate audio: {round(inference_time*1000)} milliseconds")
|
209 |
+
real_time_factor= (time.time() - t0) / out['wav'].shape[-1] * 24000
|
210 |
+
print(f"Real-time factor (RTF): {real_time_factor}")
|
211 |
+
wav_filename=f"output_{suffix}.wav"
|
212 |
+
torchaudio.save(wav_filename, torch.tensor(out["wav"]).unsqueeze(0), 24000)
|
213 |
+
return wav_filename
|
214 |
+
|
215 |
def generate_speech(history):
|
216 |
text_to_generate = history[-1][1]
|
217 |
text_to_generate = text_to_generate.replace("\n", " ").strip()
|
218 |
text_to_generate = nltk.sent_tokenize(text_to_generate)
|
|
|
|
|
|
|
|
|
219 |
|
220 |
+
language = "en"
|
|
|
|
|
221 |
|
222 |
+
wav_list = []
|
223 |
+
for i,sentence in enumerate(text_to_generate):
|
224 |
+
# Sometimes prompt </s> coming on output remove it
|
225 |
+
sentence= sentence.replace("</s>","")
|
226 |
+
# A fast fix for last chacter, may produce weird sounds if it is with text
|
227 |
+
if sentence[-1] in ["!","?",".",","]:
|
228 |
+
#just add a space
|
229 |
+
sentence = sentence[:-1] + " " + sentence[-1]
|
230 |
+
|
231 |
+
print("Sentence:", sentence)
|
232 |
+
|
233 |
+
try:
|
234 |
+
# generate speech using precomputed latents
|
235 |
+
# This is not streaming but it will be fast
|
236 |
|
237 |
+
# giving sentence suffix so we can merge all to single audio at end
|
238 |
+
# On mobile there is no autoplay support due to mobile security!
|
239 |
+
wav = get_voice(sentence,language, latent_map["Female_Voice"], suffix=i)
|
240 |
+
wav_list.append(wav)
|
241 |
+
|
242 |
+
yield wav
|
243 |
+
wait_time= librosa.get_duration(path=wav)
|
244 |
+
print("Sleeping till audio end")
|
245 |
+
time.sleep(wait_time)
|
246 |
|
247 |
except RuntimeError as e :
|
248 |
if "device-side assert" in str(e):
|
|
|
257 |
else:
|
258 |
print("RuntimeError: non device-side assert error:", str(e))
|
259 |
raise e
|
260 |
+
#Spoken on autoplay everysencen now produce a concataned one at the one
|
261 |
+
#requires pip install ffmpeg-python
|
262 |
+
files_to_concat= [ffmpeg.input(w) for w in wav_list]
|
263 |
+
combined_file_name="combined.wav"
|
264 |
+
ffmpeg.concat(*files_to_concat,v=0, a=1).output(combined_file_name).run(overwrite_output=True)
|
265 |
+
|
266 |
+
return gr.Audio.update(value=combined_file_name, autoplay=False)
|
267 |
+
|
268 |
|
269 |
with gr.Blocks(title=title) as demo:
|
270 |
gr.Markdown(DESCRIPTION)
|
|
|
288 |
btn = gr.Audio(source="microphone", type="filepath", scale=4)
|
289 |
|
290 |
with gr.Row():
|
291 |
+
audio = gr.Audio(type="numpy", streaming=False, autoplay=True, label="Generated audio response", show_label=True)
|
292 |
|
293 |
clear_btn = gr.ClearButton([chatbot, audio])
|
294 |
|
|
|
312 |
gr.Markdown("""
|
313 |
This Space demonstrates how to speak to a chatbot, based solely on open-source models.
|
314 |
It relies on 3 models:
|
315 |
+
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).
|
316 |
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).
|
317 |
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.
|
318 |
|
319 |
Note:
|
320 |
- By using this demo you agree to the terms of the Coqui Public Model License at https://coqui.ai/cpml""")
|
321 |
demo.queue()
|
322 |
+
demo.launch(debug=True)
|
requirements.txt
CHANGED
@@ -53,8 +53,11 @@ encodec==0.1.*
|
|
53 |
# deps for XTTS
|
54 |
unidecode==1.3.*
|
55 |
langid
|
56 |
-
# Install
|
57 |
-
|
|
|
58 |
deepspeed==0.8.3
|
59 |
pydub
|
|
|
|
|
60 |
gradio_client
|
|
|
53 |
# deps for XTTS
|
54 |
unidecode==1.3.*
|
55 |
langid
|
56 |
+
# Install Coqui TTS
|
57 |
+
TTS==0.17.8
|
58 |
+
# Deepspeed for fast inference
|
59 |
deepspeed==0.8.3
|
60 |
pydub
|
61 |
+
librosa
|
62 |
+
ffmpeg-python
|
63 |
gradio_client
|