tts speaks MP3: a seventh of WAV's bytes over a slow link to the server

A sentence's length for `append` comes from the Xing header LAME writes into
the first frame. The cache key carries the format, so no WAV is reused.
This commit is contained in:
Melvin Strobl
2026-10-19 21:55:17 +02:00
parent 16383ea6e2
commit f50d8b9570
4 changed files with 35 additions and 22 deletions
+20 -7
View File
@@ -10,12 +10,10 @@ from __future__ import annotations
import asyncio
import hashlib
import io
import logging
import re
import threading
import time
import wave
from collections.abc import Coroutine, Iterable
from concurrent.futures import Future
from typing import TYPE_CHECKING, Any
@@ -68,6 +66,23 @@ def sentences(text: str) -> list[str]:
return [s for s in (p.strip() for p in re.split(r"(?<=[.!?])\s+", text)) if s]
def mp3_seconds(data: bytes) -> float:
"""An MP3's length, from the frame count LAME writes into its first frame."""
header = int.from_bytes(data[:4], "big")
version = (header >> 19) & 3 # 3 is MPEG-1, 2 MPEG-2, 0 MPEG-2.5
rates = {
3: (44100, 48000, 32000),
2: (22050, 24000, 16000),
0: (11025, 12000, 8000),
}
rate = rates[version][(header >> 10) & 3]
at = max(data.find(b"Xing", 0, 64), data.find(b"Info", 0, 64))
if at < 0:
raise ValueError("the MP3 has no Xing or Info header to read its length from")
frames = int.from_bytes(data[at + 8 : at + 12], "big")
return round(frames * (1152 if version == 3 else 576) / rate, 3)
class _Speech(Node):
"""What both nodes share: the url, the artifact store and their health."""
@@ -334,19 +349,17 @@ class TtsNode(_Speech):
def _spoken(self, url: str, sentence: str) -> dict[str, Any]:
"""The reference to ``sentence`` spoken, from the cache or the server."""
key = hashlib.sha256(f"{self.cfg.voice}\n{sentence}".encode()).hexdigest()
key = hashlib.sha256(f"mp3\n{self.cfg.voice}\n{sentence}".encode()).hexdigest()
# Kept in state, so the sweep keeps the bytes it refers to.
spoken = dict(self.recall("spoken") or {})
reference = spoken.get(key)
if reference is not None and self._artifacts.has(reference["digest"]):
return dict(reference)
data = asyncio.run(client.speak(url + "/v1", sentence, self.cfg.voice))
with wave.open(io.BytesIO(data)) as clip:
seconds = round(clip.getnframes() / clip.getframerate(), 3)
stored: dict[str, Any] = self._artifacts.put(
[data], name="speech.wav", media_type="audio/wav"
[data], name="speech.mp3", media_type="audio/mpeg"
)
reference = stored | {"meta": {"seconds": seconds}}
reference = stored | {"meta": {"seconds": mp3_seconds(data)}}
spoken.pop(key, None)
spoken[key] = reference
self.remember("spoken", dict(list(spoken.items())[-SPOKEN:]))