future/smart-speaker
1
1
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

silero v5

Browse Source
This commit is contained in:
nvfuture
2026-01-07 17:31:22 +03:00
parent ebaed3fbbe
commit 7b79593cad
5 changed files with 183 additions and 87 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
ai.py
View File

@@ -13,7 +13,8 @@ SYSTEM_PROMPT = """Ты — Александр, умный голосовой а
Твоя главная цель — помогать пользователю и поддерживать интересный диалог.
Отвечай кратко и по существу, на русском языке.
Избегай длинных списков, сложного форматирования и спецсимволов, так как твои ответы озвучиваются голосом.
Пиши в разговорном стиле, как при живом общении, но не забывай о вежливости и правильности твоих ответов."""
Пиши в разговорном стиле, как при живом общении, но не забывай о вежливости и правильности твоих ответов.
ВАЖНО: Не используй в ответах панибратские или сленговые приветствия и обращения, такие как "Эй", "Хэй", "Слушай" в начале фразы и подобные."""
def ask_ai(messages_history: list) -> str:

3
cleaner.py
View File

@@ -251,6 +251,9 @@ def clean_response(text: str) -> str:
# Remove HTML tags if any
text = re.sub(r'<[^>]+>', '', text)
# Remove informal slang greetings at the beginning of sentences/responses
text = re.sub(r'^(Эй|Хэй|Слушай|Так|Ну|Короче|В\s+общем)[,!?:]?\s*', '', text, flags=re.IGNORECASE | re.MULTILINE)
# Convert numbers to words (Russian)
text = numbers_to_words(text)

2
config.py
View File

@@ -29,5 +29,5 @@ SAMPLE_RATE = 16000
CHANNELS = 1
# TTS configuration
TTS_SPEAKER = "xenia" # Available: aidar, baya, kseniya, xenia, eugene
TTS_SPEAKER = "eugene" # Available: aidar, baya, kseniya, xenia, eugene
TTS_SAMPLE_RATE = 48000

80
requirements.txt
View File

@@ -1,28 +1,52 @@
# Smart Speaker Dependencies
# Python 3.12.8
# Wake word detection
pvporcupine>=3.0.0
# Speech-to-Text
vosk>=0.3.45
# Audio
pyaudio>=0.2.14
sounddevice>=0.4.6
# AI API
requests>=2.31.0
# Environment
python-dotenv>=1.0.0
# TTS (Silero)
torch>=2.0.0
torchaudio>=2.0.0
omegaconf>=2.3.0
# Utils
numpy>=1.24.0
num2words
pymorphy3
antlr4-python3-runtime==4.9.3
certifi==2025.11.12
cffi==2.0.0
charset-normalizer==3.4.4
DAWG2-Python==0.9.0
docopt==0.6.2
filelock==3.20.1
fsspec==2025.12.0
idna==3.11
Jinja2==3.1.6
MarkupSafe==3.0.3
mpmath==1.3.0
networkx==3.6.1
num2words==0.5.14
numpy==2.4.0
nvidia-cublas-cu12==12.8.4.1
nvidia-cuda-cupti-cu12==12.8.90
nvidia-cuda-nvrtc-cu12==12.8.93
nvidia-cuda-runtime-cu12==12.8.90
nvidia-cudnn-cu12==9.10.2.21
nvidia-cufft-cu12==11.3.3.83
nvidia-cufile-cu12==1.13.1.3
nvidia-curand-cu12==10.3.9.90
nvidia-cusolver-cu12==11.7.3.90
nvidia-cusparse-cu12==12.5.8.93
nvidia-cusparselt-cu12==0.7.1
nvidia-nccl-cu12==2.27.5
nvidia-nvjitlink-cu12==12.8.93
nvidia-nvshmem-cu12==3.3.20
nvidia-nvtx-cu12==12.8.90
omegaconf==2.3.0
pvporcupine==4.0.1
PyAudio==0.2.14
pycparser==2.23
pymorphy3==2.0.6
pymorphy3-dicts-ru==2.4.417150.4580142
python-dotenv==1.2.1
PyYAML==6.0.3
requests==2.32.5
scipy==1.16.3
setuptools==80.9.0
sounddevice==0.5.3
srt==3.5.3
sympy==1.14.0
torch==2.9.1
torchaudio==2.9.1
tqdm==4.67.1
triton==3.5.1
typing_extensions==4.15.0
urllib3==2.6.2
vosk==0.3.45
websockets==15.0.1

104
tts.py
View File

@@ -3,13 +3,19 @@ Text-to-Speech module using Silero TTS.
Generates natural Russian speech with Xenia voice.
Supports interruption via wake word detection using threading.
"""
import torch
import sounddevice as sd
import numpy as np
import threading
import time
import warnings
import re
from config import TTS_SPEAKER, TTS_SAMPLE_RATE
# Suppress Silero TTS warning about text length
warnings.filterwarnings("ignore", message="Text string is longer than 1000 symbols")
class TextToSpeech:
"""Text-to-Speech using Silero TTS with wake word interruption support."""
@@ -23,17 +29,59 @@ class TextToSpeech:
def initialize(self):
"""Initialize Silero TTS model."""
print("📦 Загрузка модели Silero TTS...")
print("📦 Загрузка модели Silero TTS v5...")
# Load Silero TTS model
device = torch.device('cpu')
self.model, _ = torch.hub.load(
repo_or_dir='snakers4/silero-models',
model='silero_tts',
language='ru',
speaker='v4_ru'
repo_or_dir="snakers4/silero-models",
model="silero_tts",
language="ru",
speaker="v5_ru",
)
self.model.to(device)
print(f"✅ Модель TTS загружена (голос: {self.speaker})")
print(f"✅ Модель TTS v5 загружена (голос: {self.speaker})")
def _split_text(self, text: str, max_length: int = 900) -> list[str]:
"""Split text into chunks smaller than max_length."""
if len(text) <= max_length:
return [text]
chunks = []
# Split by sentence endings, keeping the punctuation
# pattern matches [.!?] followed by optional newlines
parts = re.split(r"([.!?]+\s*)", text)
current_chunk = ""
# Reconstruct sentences. re.split with groups returns [text, delimiter, text, delimiter...]
# We iterate through parts. If part is a delimiter (matches pattern), we append to previous text.
for part in parts:
# If the part combined with current_chunk exceeds max_length, save current_chunk
if len(current_chunk) + len(part) > max_length:
if current_chunk:
chunks.append(current_chunk.strip())
current_chunk = ""
current_chunk += part
# If even a single part is too big (very long sentence without punctuation), force split
while len(current_chunk) > max_length:
# Try to split by space
split_idx = current_chunk.rfind(" ", 0, max_length)
if split_idx == -1:
# No space found, hard cut
split_idx = max_length
chunks.append(current_chunk[:split_idx].strip())
current_chunk = current_chunk[split_idx:].lstrip()
if current_chunk:
chunks.append(current_chunk.strip())
# Filter empty chunks
return [c for c in chunks if c]
def speak(self, text: str, check_interrupt=None) -> bool:
"""
@@ -52,17 +100,28 @@ class TextToSpeech:
if not self.model:
self.initialize()
# Split text into manageable chunks
chunks = self._split_text(text)
total_chunks = len(chunks)
if total_chunks > 1:
print(f"🔊 Озвучивание (частей: {total_chunks}): {text[:50]}...")
else:
print(f"🔊 Озвучивание: {text[:50]}...")
self._interrupted = False
self._stop_flag.clear()
success = True
for i, chunk in enumerate(chunks):
if self._interrupted:
break
try:
# Generate audio
# Generate audio for chunk
audio = self.model.apply_tts(
text=text,
speaker=self.speaker,
sample_rate=self.sample_rate
text=chunk, speaker=self.speaker, sample_rate=self.sample_rate
)
# Convert to numpy array
@@ -70,16 +129,28 @@ class TextToSpeech:
if check_interrupt:
# Play with interrupt checking in parallel thread
return self._play_with_interrupt(audio_np, check_interrupt)
if not self._play_with_interrupt(audio_np, check_interrupt):
success = False
break
else:
# Standard playback
sd.play(audio_np, self.sample_rate)
sd.wait()
print("✅ Воспроизведение завершено")
return True
except Exception as e:
print(f"❌ Ошибка TTS: {e}")
print(f"❌ Ошибка TTS (часть {i + 1}/{total_chunks}): {e}")
success = False
# Continue with next chunk? or break?
# Usually if one fails, we might want to try others, but for "too long" error
# splitting should solve it. If it fails for other reasons, maybe better to stop.
# But let's keep trying subsequent chunks in case it's a specific symbol issue.
if success and not self._interrupted:
print("✅ Воспроизведение завершено")
return True
elif self._interrupted:
return False
else:
return False
def _check_interrupt_worker(self, check_interrupt):
@@ -109,9 +180,7 @@ class TextToSpeech:
"""
# Start interrupt checker thread
checker_thread = threading.Thread(
target=self._check_interrupt_worker,
args=(check_interrupt,),
daemon=True
target=self._check_interrupt_worker, args=(check_interrupt,), daemon=True
)
checker_thread.start()
@@ -133,7 +202,6 @@ class TextToSpeech:
if self._interrupted:
return False
print("✅ Воспроизведение завершено")
return True
@property