Hacker-Software/ARIA-AGENT
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases 245 Wiki Activity

feat(voxtral): Speaker-ID portiert (nur Stefans Stimme) — E3a

Browse Source 
Voxtral hatte 0 Speaker-Filter (mit Voxtral reagierte ARIA auf JEDE Stimme). Jetzt portiert aus der whisper-Bridge: speaker_id.py (ECAPA/speechbrain) kopiert, Einmal-Check auf die ersten 1.5s (fremde Stimme → leeres stt_endpoint reason=speaker_mismatch, kein Transcribe/Brain), voice_id_enroll/status/delete-RVS-Handler + voiceIdThreshold-config. voice-id-Volume gemountet, speechbrain in requirements. Ohne Enrollment fail-open.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
duffyduck
2026-08-15 13:00:32 +02:00
co-authored by Claude Opus 4.8
parent e7da9cf9c4
commit 7bc3f827d0
5 changed files with 341 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
xtts/voxtral/requirements.txt
+1
View File
@@ -3,6 +3,7 @@
transformers>=4.54
mistral-common[audio]>=1.8.1
accelerate>=0.30
speechbrain>=1.0 # Speaker-ID (ECAPA-TDNN) — nur Stefans Stimme
soundfile>=0.12
librosa>=0.10 # VoxtralProcessor.load_audio_as nutzt librosa zum WAV-Laden
numpy>=1.24