feat: XTTS auf local-Mode (dauerhaft im VRAM) + /tts_stream + Fallback

Root cause der langen Render-Zeiten und /tts_stream 400-Errors:
daswer123 default ist apiManual/api-Mode — Modell wird pro Request
gefetched/reloaded, Streaming unsupported.

Fix in xtts/docker-compose.yml:
  command: ['--listen', '-p', '8020', '-t', 'http://0.0.0.0:8020',
            '-ms', 'local',
            '-o', '/app/output', '-mf', '/app/xtts_models', '-sf', '/voices']

-ms local:
  - Modell dauerhaft im GPU-VRAM (~2GB, passt auf RTX 3060 mit 12GB)
  - Render startet sofort, kein per-Request-Load mehr
  - /tts_stream unterstuetzt → echtes progressive streaming
  - time-to-first-audio ~500ms statt 8-11s

xtts/bridge.js:
  /tts_stream primary, /tts_to_audio/ als Fallback wenn Stream fehlt.
  Robust: wenn User spaeter den Mode wieder umstellt, fallback greift.

Erste Ladung nach dem Wechsel dauert einmalig laenger (Modell ins VRAM
laden). Danach: schnell + streaming.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
duffyduck 2026-04-22 17:38:53 +02:00
parent 647a1cb726
commit 4cbe184faa
2 changed files with 27 additions and 10 deletions

View File

@ -157,16 +157,24 @@ async function _runTTSRequest(payload) {
});
};
// /tts_stream funktioniert nur bei XTTS im local-Modus. daswer123 im
// Remote-API-Modus antwortet mit 400 'HTTP Streaming is only supported
// for local models'. Wir nutzen stabil /tts_to_audio/ (batch render,
// dann Response-Body chunkweise als PCM).
// /tts_stream fuer echtes Streaming (funktioniert im XTTS local-Mode).
// Wenn Server im apiManual/api-Mode laeuft: 400 → Fallback auf /tts_to_audio/.
try {
await streamXTTSAsPCM(
cleanText,
language || "de",
hasCustomVoice ? voiceSample : null,
onChunk,
);
} catch (streamErr) {
log(`/tts_stream fehlgeschlagen (${streamErr.message.slice(0, 100)}) — Fallback /tts_to_audio/`);
await streamXTTSBatch(
cleanText,
language || "de",
hasCustomVoice ? voiceSample : null,
onChunk,
);
}
// Am Ende: final-Flag damit App weiss "fertig" und Cache geschrieben werden kann
if (pcmMeta) {

View File

@ -33,6 +33,15 @@ services:
- ./voices:/voices # Custom Voice Samples
environment:
- COQUI_TOS_AGREED=1
# Local-Modus: Modell bleibt dauerhaft im GPU-VRAM (~2GB). Vorteile:
# - Render startet sofort (kein reload pro Request)
# - /tts_stream funktioniert → echtes Streaming mit ~500ms time-to-first-audio
# Ohne diesen command: apiManual-Modus, jede Anfrage lädt Modell neu, kein Streaming.
command: ["--listen", "-p", "8020", "-t", "http://0.0.0.0:8020",
"-ms", "local",
"-o", "/app/output",
"-mf", "/app/xtts_models",
"-sf", "/voices"]
restart: unless-stopped
# ─── XTTS Bridge (verbindet zu RVS) ───────────