Remove torchcodec workaround; recommend Whisper node for ref_text

Users should connect a ComfyUI Whisper node to ref_text instead of
relying on omnivoice's internal ASR. Removes the error-catch workaround
and updates the tooltip accordingly.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

nodes/generator.py

@@ -60,7 +60,7 @@ class OmniVoiceGenerate:
                 }),
                 "ref_text": ("STRING", {
                     "default": "",
-                    "tooltip": "Transcription of ref_audio. Strongly recommended: type it manually. Auto-transcription requires FFmpeg shared libraries; if absent (e.g. some Docker images), generation will fail with a clear error message.",
+                    "tooltip": "Transcription of ref_audio. Connect a Whisper (or other STT) node for best results.",
                 }),
                 "instruct": ("STRING", {
                     "default": "",
@@ -115,16 +115,7 @@ class OmniVoiceGenerate:
                 kwargs["ref_audio"] = tmp_path
                 if ref_text:
                     kwargs["ref_text"] = ref_text
-                try:
-                    audio_tensors = model.generate(**kwargs)
-                except RuntimeError as e:
-                    if "torchcodec" in str(e).lower() or "libtorchcodec" in str(e).lower():
-                        raise RuntimeError(
-                            "Auto-transcription of the reference audio failed because FFmpeg is not "
-                            "available in this environment (required by transformers 5.x for Whisper ASR). "
-                            "Fix: type the transcript of your reference audio into the ref_text field."
-                        ) from None
-                    raise
+                audio_tensors = model.generate(**kwargs)
             finally:
                 try:
                     os.unlink(tmp_path)