alex
88ef6f11eb
- Add voice button and voice flow services for speech interaction - Add text-to-speech (TTS) service and narration - Add fullscreen button component - Update dice logic with hardened numDice fallback - Add voice integration tests - Update themes, game routes, and intent classifier - Add developer notes and QA directory
77 lines
2.5 KiB
Python
77 lines
2.5 KiB
Python
"""Consistent Danish text-to-speech for every client browser."""
|
|
|
|
import asyncio
|
|
from functools import lru_cache
|
|
import hashlib
|
|
import os
|
|
from pathlib import Path
|
|
import threading
|
|
|
|
import edge_tts
|
|
|
|
|
|
DANISH_VOICE = "da-DK-ChristelNeural"
|
|
TTS_TIMEOUT_SECONDS = 5
|
|
TTS_CACHE_DIR = Path(
|
|
os.getenv("PONY_TTS_CACHE_DIR", Path(__file__).resolve().parents[2] / "instance" / "tts_cache")
|
|
)
|
|
|
|
# Waitress has only a small worker pool. Edge TTS is a remote service, so a
|
|
# temporary outage must not be allowed to occupy every worker and freeze the
|
|
# game API. One synthesis is enough in normal play; successful MP3s are then
|
|
# served from the cache below.
|
|
_synthesis_slot = threading.BoundedSemaphore(1)
|
|
|
|
|
|
class TextToSpeechError(RuntimeError):
|
|
"""Raised when the remote speech engine cannot synthesize audio."""
|
|
|
|
|
|
async def _synthesize(text):
|
|
audio = bytearray()
|
|
communicator = edge_tts.Communicate(text, DANISH_VOICE, rate="-10%")
|
|
async for chunk in communicator.stream():
|
|
if chunk["type"] == "audio":
|
|
audio.extend(chunk["data"])
|
|
if not audio:
|
|
raise TextToSpeechError("Talegeneratoren returnerede ingen lyd")
|
|
return bytes(audio)
|
|
|
|
|
|
@lru_cache(maxsize=256)
|
|
def synthesize_danish(text):
|
|
"""Return an MP3 spoken by the fixed Danish female voice Christel."""
|
|
cache_key = hashlib.sha256(
|
|
f"{DANISH_VOICE}\0-10%\0{text}".encode("utf-8")
|
|
).hexdigest()
|
|
cache_file = TTS_CACHE_DIR / f"{cache_key}.mp3"
|
|
try:
|
|
cached = cache_file.read_bytes()
|
|
if cached:
|
|
return cached
|
|
except OSError:
|
|
pass
|
|
|
|
if not _synthesis_slot.acquire(blocking=False):
|
|
raise TextToSpeechError("Talegeneratoren er optaget")
|
|
try:
|
|
audio = asyncio.run(asyncio.wait_for(_synthesize(text), timeout=TTS_TIMEOUT_SECONDS))
|
|
try:
|
|
TTS_CACHE_DIR.mkdir(parents=True, exist_ok=True)
|
|
temporary_file = cache_file.with_suffix(".tmp")
|
|
temporary_file.write_bytes(audio)
|
|
temporary_file.replace(cache_file)
|
|
except OSError:
|
|
# The memory cache still makes this request useful when disk storage
|
|
# is temporarily unavailable.
|
|
pass
|
|
return audio
|
|
except TimeoutError as exc:
|
|
raise TextToSpeechError("Talegeneratoren svarede ikke i tide") from exc
|
|
except TextToSpeechError:
|
|
raise
|
|
except Exception as exc:
|
|
raise TextToSpeechError("Kunne ikke generere dansk tale") from exc
|
|
finally:
|
|
_synthesis_slot.release()
|