Just MVT demo for the concept of AI audio assistant.

Signed-off-by: Alex Lau (AvengerMoJo) <alau@suse.com>
This commit is contained in:
Alex Lau 2023-11-10 02:26:16 +08:00
parent 35055a80b4
commit 7ba9d8d3db
No known key found for this signature in database
GPG Key ID: 3F28752074CC2261
11 changed files with 569 additions and 0 deletions

51
audio_utils.py Normal file
View File

@ -0,0 +1,51 @@
# audio_utils.py
import csv
import math
import os
from pydub import AudioSegment
from tqdm import tqdm
import tempfile
class AudioSplit():
"""Export audio file into smaller size and format that AI can understand."""
def __init__(self, filename, title="Output"):
"""Init the Audio utils."""
self.folder = tempfile.TemporaryDirectory()
self.title = os.path.splitext(title)[0]
os.mkdir(self.folder.name + '/audio/')
if filename.endswith('.mp3'):
another = tempfile.TemporaryDirectory()
sound = AudioSegment.from_mp3(filename)
sound = sound.set_frame_rate(16000)
sound.export(another.name + '/orginal.wav', format="wav")
self.audio = AudioSegment.from_wav(another.name + '/orginal.wav')
elif filename.endswith('.wav'):
self.audio = AudioSegment.from_wav(filename)
self.metadata = 'metadata.csv'
def get_folder(self):
return self.folder.name
def get_duration(self):
return self.audio.duration_seconds
def single_split(self, from_sec, to_sec, split_filename):
time1 = from_sec * 1000
time2 = to_sec * 1000
split_audio = self.audio[time1:time2]
split_audio.export(self.folder.name + '/audio/' + split_filename, format="wav")
def multiple_split(self, sec_per_split):
with open(self.folder.name + '/audio/' + self.metadata, 'w', encoding='utf8') as csv_file:
writer = csv.writer(csv_file)
header = ['id', 'file_name', 'transcription']
writer.writerow(header)
total_sec = math.ceil(self.get_duration() / sec_per_split)
for i in tqdm(range(0, total_sec)):
split_fn = self.title + '_' + str(i) + '.wav'
self.single_split(i*sec_per_split, i*sec_per_split + sec_per_split, split_fn)
data = [i, split_fn, "Export wave file " + str(i)]
writer.writerow(data)

BIN
data/audio/i_dont_get.wav Normal file

Binary file not shown.

BIN
data/audio/suse_intro.wav Normal file

Binary file not shown.

163
demo_assistant.py Normal file
View File

@ -0,0 +1,163 @@
import io
import tempfile
import pyaudio
from pydub import AudioSegment
import wave
import re
import time
import queue
import threading
from transformers import pipeline
from datasets import load_dataset
from faster_whisper import WhisperModel
import torch
from TTS.api import TTS
from gpt4all import GPT4All
from audio_utils import AudioSplit
CHUNK = 1024
FORMAT = pyaudio.paInt16 # 16-bit resolution
CHANNELS = 1
RATE = 16000 # sample rate
DURATION = 3
Q_DURATION = 7
SUSE = r"s*u*s*e"
THANK= r"Thank\s*(?:you|u)\b"
g_active = False
g_wait = False
g_lock = threading.Lock()
counter = 0
p_audio = pyaudio.PyAudio()
playback_stream = p_audio.open(format=FORMAT, channels=CHANNELS, rate=24000, output=True)
record_stream = p_audio.open(format=FORMAT, channels=CHANNELS, rate=RATE, input=True, frames_per_buffer=CHUNK)
def record_audio():
global p_audio
global g_active
global g_wait
print("Recording started...")
while True:
if g_active == True:
print("c", end="")
if g_wait == True:
print("w", end="")
while not audio_queue.empty():
audio_queue.get()
time.sleep(1)
continue
audio_data = record_stream.read(CHUNK, exception_on_overflow=False)
audio_queue.put(audio_data)
print(".", end="")
record_stream.stop_stream()
record_stream.close()
p_audio.terminate()
def speech_to_text():
global p_audio
global g_active
global g_wait
while True:
if g_active:
time_duration = Q_DURATION
else:
time_duration = DURATION
tf = tempfile.NamedTemporaryFile(suffix=".wav", delete=True, mode='wb')
mp3_tf = tempfile.NamedTemporaryFile(suffix=".mp3", delete=True, mode='wb')
with wave.open(tf.name, 'wb') as wav_file:
wav_file.setnchannels(CHANNELS)
wav_file.setsampwidth(p_audio.get_sample_size(FORMAT))
wav_file.setframerate(RATE)
# Read audio data from the stream for the specified duration
for i in range(0, RATE // CHUNK * time_duration):
print("r", end="")
audio_data = audio_queue.get()
wav_file.writeframes(audio_data)
audio_queue.task_done()
print(f"{time_duration} sec recording done.")
# Perform speech recognition
audio = AudioSegment.from_wav(tf.name)
audio.export(mp3_tf.name, format="mp3")
# segments, info = model.transcribe(mp3_tf_filename, beam_size=5)
segments, _ = model.transcribe(mp3_tf.name)
questions = []
if g_active:
counter += 1
for segment in segments:
print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))
if segment.text:
questions.append(segment.text)
q = re.sub(THANK, "", " ".join(questions))
print(f"Question:{q} counter{counter}")
if len(q) > 40 and counter >= 2:
counter = 0
output = gpt_model.generate(" ".join(questions), max_tokens=50)
print(f"Answer:{output}")
reply_wav = tempfile.NamedTemporaryFile(suffix=".wav", delete=True, mode='wb')
tts.tts_to_file(text=output, file_path=reply_wav.name)
play(playback_stream, reply_wav.name)
with g_lock:
g_active = False
time.sleep(5)
continue
for segment in segments:
print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))
text_input = segment.text.lower()
if text_input.find("hey") != -1:
if re.search(SUSE, text_input):
counter = 1
with g_lock:
g_active = True
g_wait = True
play(playback_stream, "data/audio/suse_intro.wav")
print("Finish suse")
with g_lock:
g_wait = False
time.sleep(5)
def play(play_stream, filename):
wave_file = wave.open(filename, 'rb')
print(f"Wave: rate={wave_file.getframerate()} channels={wave_file.getnchannels()} width={wave_file.getsampwidth()}")
out_data = wave_file.readframes(CHUNK)
while out_data:
play_stream.write(out_data)
out_data = wave_file.readframes(CHUNK)
# Get device
device = "cuda" if torch.cuda.is_available() else "cpu"
# Init TTS
# tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to(device)
tts = TTS("tts_models/en/blizzard2013/capacitron-t2-c150_v2").to(device)
gpt_model = GPT4All("orca-mini-3b-gguf2-q4_0.gguf")
# Create a queue to share audio data between threads
audio_queue = queue.Queue()
# model_size = "large-v2"
model_size = "small.en"
# model_size = "tiny.en"
# Run on GPU with FP16
model = WhisperModel(model_size, device="cpu", compute_type="int8")
print(f"Stream: playback->{playback_stream.get_write_available()}")
# Create and start the recording thread
recording_thread = threading.Thread(target=record_audio)
recording_thread.start()
# Create and start the speech-to-text thread
speech_to_text_thread = threading.Thread(target=speech_to_text)
speech_to_text_thread.start()
# Wait for the recording thread to finish (you can define conditions to stop the recording)
recording_thread.join()
# Stop the speech-to-text thread
speech_to_text_thread.join()
p.terminate()

8
requirements.txt Normal file
View File

@ -0,0 +1,8 @@
pyaudio
pydub
transformers
datasets
faster_whisper
torch
TTS
gpt4all

1
setup.sh Normal file
View File

@ -0,0 +1 @@
sudo zypper install python311-devel portaudio-devel espeak-ng

25
test/coqui_tts_out.py Normal file
View File

@ -0,0 +1,25 @@
import io
import tempfile
import pyaudio
from pydub import AudioSegment
import wave
from transformers import pipeline
from datasets import load_dataset
from audio_utils import AudioSplit
from faster_whisper import WhisperModel
import torch
from TTS.api import TTS
# Get device
device = "cuda" if torch.cuda.is_available() else "cpu"
# Init TTS
# tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to(device)
tts = TTS("tts_models/en/blizzard2013/capacitron-t2-c150_v2").to(device)
# tts.tts_to_file(text="Hello, I am your SUSE ... assistant. What can I do for you today?", file_path="/tmp/out.wav")
tts.tts_to_file(text="Sorry, I don't get your question, please say that again please.", file_path="/tmp/out.wav")

111
test/fast_whisper.py Normal file
View File

@ -0,0 +1,111 @@
import io
import tempfile
import pyaudio
from pydub import AudioSegment
import wave
import re
import queue
from transformers import pipeline
from datasets import load_dataset
from faster_whisper import WhisperModel
import torch
from TTS.api import TTS
from audio_utils import AudioSplit
CHUNK = 1024
FORMAT = pyaudio.paInt16 # 16-bit resolution
CHANNELS = 1
RATE = 16000 # sample rate
DURATION = 2
SUSE = r"s*u*s*e"
p = pyaudio.PyAudio()
def record(stream):
print("Recording started...")
while True:
audio_data = stream.read(CHUNK)
audio_queue.put(audio_data)
stream.stop_stream()
stream.close()
audio.terminate()
def play(play_stream, filename):
wave_file = wave.open(filename, 'rb')
print(f"Wave: rate={wave_file.getframerate()} channels={wave_file.getnchannels()} width={wave_file.getsampwidth()}")
out_data = wave_file.readframes(CHUNK)
while out_data:
play_stream.write(out_data)
out_data = wave_file.readframes(CHUNK)
# Get device
device = "cuda" if torch.cuda.is_available() else "cpu"
# Init TTS
# tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to(device)
tts = TTS("tts_models/en/blizzard2013/capacitron-t2-c150_v2").to(device)
# Create a queue to share audio data between threads
audio_queue = queue.Queue()
# model_size = "large-v2"
model_size = "small.en"
# model_size = "tiny.en"
# Run on GPU with FP16
model = WhisperModel(model_size, device="cpu", compute_type="int8")
device = ""
for i in range(p.get_device_count()):
device = p.get_device_info_by_index(i)
if device['name']=="default":
print(device)
break
playback_stream = p.open(format=p.get_format_from_width(2),
channels=1,
rate=24000,
# output_device_index = device['index'],
output=True)
print(f"Stream: playback->{playback_stream.get_write_available()}")
# generator = pipeline(task="automatic-speech-recognition", model="microsoft/speecht5_asr")
while True:
tf = tempfile.NamedTemporaryFile(suffix=".wav", delete=True, mode='wb')
mp3_tf = tempfile.NamedTemporaryFile(suffix=".mp3", delete=True, mode='wb')
temp_filename = tf.name
mp3_tf_filename = mp3_tf.name
with wave.open(temp_filename, 'wb') as wav_file:
wav_file.setnchannels(CHANNELS)
wav_file.setsampwidth(p.get_sample_size(FORMAT))
wav_file.setframerate(RATE)
stream = p.open(format=FORMAT, channels=CHANNELS, rate=RATE, input=True, frames_per_buffer=CHUNK)
print("Listening...")
frames = []
for i in range(0, RATE // CHUNK * DURATION):
# Read audio data from the stream for the specified duration
audio_data = stream.read(CHUNK)
frames.append(audio_data)
wav_file.writeframes(audio_data)
#print(f"{DURATION} sec recording done.")
stream.close()
audio = AudioSegment.from_wav(temp_filename)
audio.export(mp3_tf_filename, format="mp3")
# segments, info = model.transcribe(mp3_tf_filename, beam_size=5)
# print("Detected language '%s' with probability %f" % (info.language, info.language_probability))
segments, _ = model.transcribe(mp3_tf_filename)
for segment in segments:
print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))
# out_wav = tempfile.NamedTemporaryFile(suffix=".mp3", delete=True, mode='wb')
text_input = segment.text.lower()
if text_input.find("hey") != -1:
if re.search(SUSE, text_input):
# wav = tts.tts(text=segment.text, speaker_wav=speak_wav.name, language="en")
# tts.tts_to_file(text="This is SUSE assistant what can I do for you today?", language="en", file_path=out_wav.name)
# tts.tts_to_file(text="This is SUSE assistant what can I do for you today?", file_path=out_wav.name)
# wave_file = wave.open(out_wav.name, 'rb')
play(playback_stream, "data/audio/suse_intro.wav")
p.terminate()

158
test/fast_whisper2.py Normal file
View File

@ -0,0 +1,158 @@
import io
import tempfile
import pyaudio
from pydub import AudioSegment
import wave
import re
import time
import queue
import threading
from transformers import pipeline
from datasets import load_dataset
from faster_whisper import WhisperModel
import torch
from TTS.api import TTS
from gpt4all import GPT4All
from audio_utils import AudioSplit
CHUNK = 1024
FORMAT = pyaudio.paInt16 # 16-bit resolution
CHANNELS = 1
RATE = 16000 # sample rate
DURATION = 5
SUSE = r"s*u*s*e"
THANK= r"Thank\s*(?:you|u)\b"
g_active = False
g_wait = False
g_lock = threading.Lock()
counter = 0
p_audio = pyaudio.PyAudio()
playback_stream = p_audio.open(format=FORMAT, channels=CHANNELS, rate=24000, output=True)
record_stream = p_audio.open(format=FORMAT, channels=CHANNELS, rate=RATE, input=True, frames_per_buffer=CHUNK)
def record_audio():
global p_audio
global g_active
global g_wait
print("Recording started...")
while True:
if g_active == True:
print("c", end="")
if g_wait == True:
print("w", end="")
while not audio_queue.empty():
audio_queue.get()
time.sleep(1)
continue
audio_data = record_stream.read(CHUNK)
audio_queue.put(audio_data)
print(".", end="")
record_stream.stop_stream()
record_stream.close()
p_audio.terminate()
def speech_to_text():
global p_audio
global g_active
global g_wait
while True:
tf = tempfile.NamedTemporaryFile(suffix=".wav", delete=True, mode='wb')
mp3_tf = tempfile.NamedTemporaryFile(suffix=".mp3", delete=True, mode='wb')
with wave.open(tf.name, 'wb') as wav_file:
wav_file.setnchannels(CHANNELS)
wav_file.setsampwidth(p_audio.get_sample_size(FORMAT))
wav_file.setframerate(RATE)
# Read audio data from the stream for the specified duration
for i in range(0, RATE // CHUNK * DURATION):
print("r", end="")
audio_data = audio_queue.get()
wav_file.writeframes(audio_data)
audio_queue.task_done()
print(f"{DURATION} sec recording done.")
# Perform speech recognition
audio = AudioSegment.from_wav(tf.name)
audio.export(mp3_tf.name, format="mp3")
# segments, info = model.transcribe(mp3_tf_filename, beam_size=5)
segments, _ = model.transcribe(mp3_tf.name)
questions = []
if g_active:
counter += 1
for segment in segments:
print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))
if segment.text:
questions.append(segment.text)
q = re.sub(THANK, "", " ".join(questions))
print(f"Question:{q} counter{counter}")
if len(q) > 40 and counter > 3:
counter = 0
output = gpt_model.generate(" ".join(questions), max_tokens=50)
print(f"Answer:{output}")
reply_wav = tempfile.NamedTemporaryFile(suffix=".wav", delete=True, mode='wb')
tts.tts_to_file(text=output, file_path=reply_wav.name)
play(playback_stream, reply_wav.name)
with g_lock:
g_active = False
time.sleep(5)
continue
for segment in segments:
print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))
text_input = segment.text.lower()
if text_input.find("hey") != -1:
if re.search(SUSE, text_input):
counter = 1
with g_lock:
g_active = True
g_wait = True
play(playback_stream, "data/audio/suse_intro.wav")
print("Finish suse")
with g_lock:
g_wait = False
time.sleep(5)
def play(play_stream, filename):
wave_file = wave.open(filename, 'rb')
print(f"Wave: rate={wave_file.getframerate()} channels={wave_file.getnchannels()} width={wave_file.getsampwidth()}")
out_data = wave_file.readframes(CHUNK)
while out_data:
play_stream.write(out_data)
out_data = wave_file.readframes(CHUNK)
# Get device
device = "cuda" if torch.cuda.is_available() else "cpu"
# Init TTS
# tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to(device)
tts = TTS("tts_models/en/blizzard2013/capacitron-t2-c150_v2").to(device)
gpt_model = GPT4All("orca-mini-3b-gguf2-q4_0.gguf")
# Create a queue to share audio data between threads
audio_queue = queue.Queue()
# model_size = "large-v2"
model_size = "small.en"
# model_size = "tiny.en"
# Run on GPU with FP16
model = WhisperModel(model_size, device="cpu", compute_type="int8")
print(f"Stream: playback->{playback_stream.get_write_available()}")
# Create and start the recording thread
recording_thread = threading.Thread(target=record_audio)
recording_thread.start()
# Create and start the speech-to-text thread
speech_to_text_thread = threading.Thread(target=speech_to_text)
speech_to_text_thread.start()
# Wait for the recording thread to finish (you can define conditions to stop the recording)
recording_thread.join()
# Stop the speech-to-text thread
speech_to_text_thread.join()
p.terminate()

19
test/pyttsx_test.py Normal file
View File

@ -0,0 +1,19 @@
import pyttsx3
def text_to_speech(text):
# Initialize the TTS engine
engine = pyttsx3.init()
# Set properties (optional)
engine.setProperty('rate', 130) # Speed of speech
engine.setProperty('volume', 0.6) # Volume level (0.0 to 1.0)
# Convert text to speech
engine.say(text)
# Wait for the speech to finish
engine.runAndWait()
# Example usage
text = "Hello, this is a simple text-to-speech example in Python."
text_to_speech(text)

33
test/test_audio.py Normal file
View File

@ -0,0 +1,33 @@
import pyaudio
import wave
filename = '/tmp/out.wav'
# Set chunk size of 1024 samples per data frame
chunk = 1024
# Open the sound file
wf = wave.open(filename, 'rb')
# Create an interface to PortAudio
p = pyaudio.PyAudio()
# Open a .Stream object to write the WAV file to
# 'output = True' indicates that the sound will be played rather than recorded
stream = p.open(format = p.get_format_from_width(wf.getsampwidth()),
channels = wf.getnchannels(),
rate = wf.getframerate(),
output = True)
# Read data in chunks
data = wf.readframes(chunk)
# Play the sound by writing the audio data to the stream
while data != '':
stream.write(data)
data = wf.readframes(chunk)
# Close and terminate the stream
stream.close()
p.terminate()