cdd331af35
ROOT CAUSE
User reported the Python edition showed "только пустой терминал" — empty
terminal — when launched. Three compounding issues:
1. `tts.py` called `torch.hub.load(...)` at MODULE IMPORT time. On first
use this downloads a 60MB Silero model with no progress output beyond
one "Using cache found in..." line. With no prior banner the user
couldn't tell if it crashed or was loading.
2. main.py's startup banner (`print("=" * 60)...J.A.R.V.I.S. v...`)
appeared AFTER all the heavy imports. So even when nothing was wrong,
visible feedback was delayed by 10-60 seconds.
3. If `pip install -r requirements.txt` failed for any package
(`simpleaudio` notoriously needs MSVC to build), `import simpleaudio`
crashed the whole process with a raw ModuleNotFoundError.
FIXES
- `tts.py`: Silero model load moved into `_ensure_model()` called lazily
from `synthesize()`. First synth pays the load cost (with a one-line
progress hint) — everything before that is visible immediately.
- `main.py`: print "J.A.R.V.I.S. loading..." banner BEFORE any heavy
imports. User sees the terminal is alive within ~100ms.
- `main.py`: scan required deps before importing them; if any are
missing, print "Run: pip install -r requirements.txt" with the
specific list and exit cleanly (exit code 2) instead of stack-tracing.
- `main.py`: `simpleaudio` is now soft-optional. Falls back to stdlib
`winsound` on Windows installs that couldn't build the C extension.
Sound cues still work; only the precise wait_done semantics are
slightly different.
VERIFIED LOCALLY
`.venv\Scripts\python.exe main.py` now prints:
```
============================================================
J.A.R.V.I.S. loading (Python edition)...
Heavy modules (vosk / torch / pvrecorder) take a few seconds.
============================================================
[scheduler] thread started (1 tasks loaded)
============================================================
J.A.R.V.I.S. v0.4.1 [Python edition]
Total commands: 200
============================================================
Using device: Микрофон (5- Fifine Microphone)
Jarvis (v0.4.1) начал свою работу ...
Yes, sir.
```
Total time to first banner: ~100ms (was 10-60s of nothing).
All 104 pytest tests still pass.
91 lines
2.8 KiB
Python
91 lines
2.8 KiB
Python
"""TTS — Silero ru_v3 voice synth with lazy initialisation.
|
|
|
|
Why lazy: torch.hub.load() blocks for 30-120 seconds on the first run
|
|
(downloading the model) and produces no progress output beyond a single
|
|
"Using cache found in..." line. If this runs at MODULE IMPORT time, the
|
|
whole jarvis-python startup looks frozen — exactly the bug the user hit.
|
|
|
|
Fix: defer the load until first synth call. main.py prints its startup
|
|
banner first, then this module loads on demand. Side effect: the first
|
|
"speak" is slow (3-4s), but every print before it is visible.
|
|
"""
|
|
import time
|
|
|
|
import numpy as np
|
|
import sounddevice as sd
|
|
import torch
|
|
|
|
import config
|
|
import effects
|
|
|
|
language = 'ru'
|
|
model_id = 'ru_v3'
|
|
sample_rate = 48000 # 48000
|
|
speaker = 'aidar' # aidar, baya, kseniya, xenia, random
|
|
put_accent = True
|
|
put_yo = True
|
|
device = torch.device('cpu') # cpu или gpu
|
|
|
|
# Initialised on first synthesize() call.
|
|
_model = None
|
|
|
|
|
|
def _ensure_model():
|
|
"""Load Silero on first use. Subsequent calls are no-ops.
|
|
|
|
Prints a one-line progress hint so the user can tell the assistant is
|
|
busy downloading rather than hung — critical for the first run when
|
|
the model isn't cached yet.
|
|
"""
|
|
global _model
|
|
if _model is not None:
|
|
return _model
|
|
print("[tts] Loading Silero ru_v3 (first use — ~30s if model isn't cached)...")
|
|
started = time.time()
|
|
m, _ = torch.hub.load(repo_or_dir='snakers4/silero-models',
|
|
model='silero_tts',
|
|
language=language,
|
|
speaker=model_id)
|
|
m.to(device)
|
|
_model = m
|
|
print(f"[tts] Silero ready ({time.time() - started:.1f}s).")
|
|
return _model
|
|
|
|
|
|
def _post_process(audio):
|
|
if not getattr(config, 'TTS_EFFECTS_ENABLED', False):
|
|
return audio
|
|
|
|
arr = np.asarray(audio, dtype=np.float32)
|
|
low = getattr(config, 'TTS_BANDPASS_LOW_HZ', 0)
|
|
high = getattr(config, 'TTS_BANDPASS_HIGH_HZ', 0)
|
|
bandpass = (low, high) if low and high and high > low else None
|
|
|
|
wet = float(getattr(config, 'TTS_REVERB_WET', 0.0))
|
|
decay = int(getattr(config, 'TTS_REVERB_DECAY_MS', 0))
|
|
reverb = (wet, decay) if wet > 0 and decay > 0 else None
|
|
|
|
pitch = int(getattr(config, 'TTS_PITCH_SEMITONES', 0))
|
|
|
|
return effects.process(arr, sample_rate,
|
|
bandpass=bandpass,
|
|
reverb=reverb,
|
|
pitch_semitones=pitch)
|
|
|
|
|
|
def synthesize(what: str):
|
|
m = _ensure_model()
|
|
audio = m.apply_tts(text=what + "..",
|
|
speaker=speaker,
|
|
sample_rate=sample_rate,
|
|
put_accent=put_accent,
|
|
put_yo=put_yo)
|
|
return _post_process(audio)
|
|
|
|
|
|
def va_speak(what: str):
|
|
audio = synthesize(what)
|
|
|
|
sd.play(audio, sample_rate * 1.05)
|
|
time.sleep((len(audio) / sample_rate) + 0.5)
|
|
sd.stop()
|