Release NVIDIA WakeWord Trainer v26

VERSION

@@ -1 +1 @@
-25
+26

WHATS_NEW.md

@@ -1,2 +1 @@
-- Prevented unreadable audio files from freezing sample preparation by adding a bounded FFmpeg watchdog, safe cleanup, and visible normalization progress.
+- Restored MOSS-TTS-Nano generation with the current voice-cloning API so all four providers contribute to the final training corpus.
-- Kept primary and guided STT results visible after captures are automatically sorted into positive or negative training samples.

cli/tts_generate_samples.py

@@ -44,7 +44,7 @@ from tts_config import (  # noqa: E402
 )
 
 
-GENERATOR_VERSION = "modern-tts-v15-four-provider-direct-corpus-safe-limits"
+GENERATOR_VERSION = "modern-tts-v16-four-provider-direct-corpus-safe-limits"
 VOICE_BANK_VERSION = "modern-tts-voice-bank-v1-native-random-qualified-single-utterance"
 COMPATIBLE_VOICE_BANK_VERSIONS = {
     VOICE_BANK_VERSION,

cli/tts_moss_worker.py

@@ -69,7 +69,6 @@ def main() -> int:
             text=str(item["text"]),
             output_audio_path=str(output_path),
             mode="voice_clone",
-            prompt_text=str(item["ref_text"]),
             prompt_audio_path=str(item["ref_audio"]),
             reference_audio_path=None,
             text_tokenizer_path=None,

tests/test_modern_tts.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import argparse
+import ast
 import importlib.util
 import json
 import math
@@ -70,6 +71,22 @@ class ModernTtsTests(unittest.TestCase):
             ["--position_temperature", "5.0", "--class_temperature", "0.0"],
         )
 
+    def test_moss_voice_clone_uses_audio_without_disallowed_prompt_text(self) -> None:
+        worker_path = REPO_ROOT / "cli" / "tts_moss_worker.py"
+        tree = ast.parse(worker_path.read_text(encoding="utf-8"))
+        inference_calls = [
+            node
+            for node in ast.walk(tree)
+            if isinstance(node, ast.Call)
+            and isinstance(node.func, ast.Attribute)
+            and node.func.attr == "inference"
+        ]
+
+        self.assertEqual(len(inference_calls), 1)
+        keywords = {keyword.arg for keyword in inference_calls[0].keywords}
+        self.assertIn("prompt_audio_path", keywords)
+        self.assertNotIn("prompt_text", keywords)
+
     def test_omnivoice_uses_a_hidden_stable_prompt_before_short_clone(self) -> None:
         with tempfile.TemporaryDirectory() as temp_dir:
             data_dir = Path(temp_dir)