chore: bump version to 1.0.6

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

.github/workflows/publish.yml

@@ -1,9 +1,12 @@
 name: Publish to ComfyUI Registry
 
 on:
+  workflow_dispatch:
   push:
-    tags:
+    branches:
-      - "v*"
+      - master
+    paths:
+      - "pyproject.toml"
 
 jobs:
   publish:

__init__.py

@@ -1,4 +1,4 @@
-from .nodes import OmniVoiceModelLoader, OmniVoiceGenerate, OmniVoiceEpubLoader, OmniVoiceVoicePreset, OmniVoiceMixVoices
+from .nodes import OmniVoiceModelLoader, OmniVoiceGenerate, OmniVoiceEpubLoader, OmniVoiceVoicePreset, OmniVoiceMixVoices, OmniVoiceVoiceDesign, OmniVoiceSpeaker, OmniVoiceSpeakers
 
 NODE_CLASS_MAPPINGS = {
     "OmniVoiceModelLoader": OmniVoiceModelLoader,
@@ -6,6 +6,9 @@ NODE_CLASS_MAPPINGS = {
     "OmniVoiceEpubLoader": OmniVoiceEpubLoader,
     "OmniVoiceVoicePreset": OmniVoiceVoicePreset,
     "OmniVoiceMixVoices": OmniVoiceMixVoices,
+    "OmniVoiceVoiceDesign": OmniVoiceVoiceDesign,
+    "OmniVoiceSpeaker": OmniVoiceSpeaker,
+    "OmniVoiceSpeakers": OmniVoiceSpeakers,
 }
 
 NODE_DISPLAY_NAME_MAPPINGS = {
@@ -14,6 +17,11 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "OmniVoiceEpubLoader": "OmniVoice EPUB Loader",
     "OmniVoiceVoicePreset": "OmniVoice Voice Preset",
     "OmniVoiceMixVoices": "OmniVoice Mix Voices",
+    "OmniVoiceVoiceDesign": "OmniVoice Voice Design",
+    "OmniVoiceSpeaker": "OmniVoice Speaker",
+    "OmniVoiceSpeakers": "OmniVoice Speakers",
 }
 
-__all__ = ["NODE_CLASS_MAPPINGS", "NODE_DISPLAY_NAME_MAPPINGS"]
+WEB_DIRECTORY = "./web"
+
+__all__ = ["NODE_CLASS_MAPPINGS", "NODE_DISPLAY_NAME_MAPPINGS", "WEB_DIRECTORY"]

nodes/__init__.py

@@ -3,5 +3,7 @@ from .generator import OmniVoiceGenerate
 from .epub_loader import OmniVoiceEpubLoader
 from .voice_presets import OmniVoiceVoicePreset
 from .mix_voices import OmniVoiceMixVoices
+from .voice_design import OmniVoiceVoiceDesign
+from .multi_speaker import OmniVoiceSpeaker, OmniVoiceSpeakers
 
-__all__ = ["OmniVoiceModelLoader", "OmniVoiceGenerate", "OmniVoiceEpubLoader", "OmniVoiceVoicePreset", "OmniVoiceMixVoices"]
+__all__ = ["OmniVoiceModelLoader", "OmniVoiceGenerate", "OmniVoiceEpubLoader", "OmniVoiceVoicePreset", "OmniVoiceMixVoices", "OmniVoiceVoiceDesign", "OmniVoiceSpeaker", "OmniVoiceSpeakers"]

nodes/epub_loader.py

@@ -58,6 +58,12 @@ def _extract_chapters(epub_path):
                     if tag:
                         title = tag.get_text(strip=True)
                         break
+            # Remove title/heading elements so they don't appear in the body text
+            if soup.title:
+                soup.title.decompose()
+            for hn in ['h1', 'h2', 'h3']:
+                for tag in soup.find_all(hn):
+                    tag.decompose()
             for tag in soup.find_all(_BLOCK_TAGS):
                 tag.append(soup.new_string('\n\n'))
             text = soup.get_text(separator='')
@@ -90,8 +96,8 @@ class OmniVoiceEpubLoader:
             },
         }
 
-    RETURN_TYPES = ("STRING", "STRING")
+    RETURN_TYPES = ("STRING", "STRING", "STRING")
-    RETURN_NAMES = ("text", "chapter_list")
+    RETURN_NAMES = ("text", "chapter_title", "chapter_list")
     FUNCTION = "load_epub"
     CATEGORY = "OmniVoice"
 
@@ -100,7 +106,7 @@ class OmniVoiceEpubLoader:
         n = len(chapters)
 
         if n == 0:
-            return ("", "")
+            return ("", "", "")
 
         start = max(1, min(chapter_start, n))
         end   = max(start, min(chapter_end, n))
@@ -111,8 +117,12 @@ class OmniVoiceEpubLoader:
             for i, ch in enumerate(chapters, 1)
         )
 
+        # chapter_title: title of the first selected chapter (useful for file naming)
+        first = chapters[start - 1]
+        chapter_title = first["title"] if first["title"] else f"Chapter {start}"
+
         # text: selected range joined by delimiter
         selected = chapters[start - 1 : end]
         text = "\n\n---\n\n".join(ch["text"] for ch in selected)
 
-        return (text, chapter_list)
+        return (text, chapter_title, chapter_list)

nodes/generator.py

@@ -1,8 +1,26 @@
+import re
 import tempfile
 import os
 import torch
 import soundfile as sf
 
+_TAG_RE = re.compile(r'^\[([^\]]+)\]:?\s*(.*)', re.DOTALL)
+
+
+def _write_tmp_wav(ref_audio):
+    """Write a ComfyUI AUDIO dict to a temp WAV file. Returns the path (caller must delete)."""
+    tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
+    tmp_path = tmp.name
+    tmp.close()
+    waveform = ref_audio["waveform"].squeeze(0).cpu()  # (channels, samples)
+    audio_np = waveform.numpy()
+    sf.write(
+        tmp_path,
+        audio_np[0] if audio_np.shape[0] == 1 else audio_np.T,
+        int(ref_audio["sample_rate"]),
+    )
+    return tmp_path
+
 
 class OmniVoiceGenerate:
     @classmethod
@@ -49,12 +67,21 @@ class OmniVoiceGenerate:
                         "tooltip": (
                             "voice_cloning  – clone the voice from ref_audio (requires ref_audio)\n"
                             "voice_design   – describe a voice with the instruct field (requires instruct)\n"
-                            "auto_voice     – model picks a voice automatically"
+                            "auto_voice     – model picks a voice automatically\n"
+                            "\n"
+                            "Ignored when a Speakers roster is connected."
                         ),
                     },
                 ),
             },
             "optional": {
+                "speakers": ("OMNIVOICE_SPEAKERS", {
+                    "tooltip": (
+                        "Connect an OmniVoice Speakers node to enable multi-speaker generation.\n"
+                        "When connected, ref_audio / instruct / mode are ignored and each paragraph\n"
+                        "is routed to its assigned speaker automatically."
+                    ),
+                }),
                 "ref_audio": ("AUDIO", {
                     "tooltip": "Reference audio clip to clone the voice from. Used in voice_cloning mode.",
                 }),
@@ -65,23 +92,32 @@ class OmniVoiceGenerate:
                 "instruct": ("STRING", {
                     "default": "",
                     "tooltip": (
-                        "Voice description for voice_design mode. Combine attributes freely.\n"
+                        "Voice style description. Required for voice_design mode; optional in voice_cloning\n"
+                        "mode to attempt accent/style transfer on top of the cloned voice.\n"
+                        "Connect the OmniVoice Voice Design node for structured input.\n"
                         "\n"
                         "GENDER:   male, female\n"
                         "AGE:      child, teenager, young adult, middle-aged, elderly\n"
-                        "PITCH:    very low, low, moderate, high, very high\n"
+                        "PITCH:    very low pitch, low pitch, moderate pitch, high pitch, very high pitch, whisper\n"
-                        "STYLE:    whisper\n"
                         "\n"
-                        "ENGLISH ACCENTS (text must be English):\n"
+                        "ACCENTS (only these are supported by the model):\n"
-                        "  american, british, australian, canadian,\n"
+                        "  american accent, australian accent, british accent, canadian accent,\n"
-                        "  indian, chinese, korean, japanese, portuguese, russian\n"
+                        "  chinese accent, indian accent, japanese accent, korean accent,\n"
+                        "  portuguese accent, russian accent\n"
                         "\n"
                         "EXAMPLE:  female, high pitch, british accent"
                     ),
                 }),
+                "guidance_scale": ("FLOAT", {
+                    "default": 2.0, "min": 0.0, "max": 20.0, "step": 0.1,
+                    "tooltip": (
+                        "Classifier-free guidance scale. Higher = more faithful to the reference/instruct, "
+                        "but can over-saturate. 2.0 is a good default."
+                    ),
+                }),
                 "speed": ("FLOAT", {
-                    "default": 1.0, "min": 0.1, "max": 3.0, "step": 0.1,
+                    "default": 1.0, "min": 0.3, "max": 3.0, "step": 0.1,
-                    "tooltip": "Playback speed multiplier. 1.0 = normal, >1.0 = faster, <1.0 = slower.",
+                    "tooltip": "Playback speed multiplier. 1.0 = normal, >1.0 = faster, <1.0 = slower. Below 0.3 produces noise and extreme VRAM usage.",
                 }),
                 "num_step": ("INT", {
                     "default": 32, "min": 1, "max": 100,
@@ -104,10 +140,17 @@ class OmniVoiceGenerate:
     FUNCTION = "generate"
     CATEGORY = "OmniVoice"
 
-    def generate(self, model, text, mode, ref_audio=None, ref_text="", instruct="", speed=1.0, num_step=32, seed=0):
+    def generate(self, model, text, mode, speakers=None, ref_audio=None, ref_text="",
+                 instruct="", guidance_scale=2.0, speed=1.0, num_step=32, seed=0):
         if seed != 0:
             torch.manual_seed(seed)
-        kwargs = {"text": text, "speed": speed, "num_step": num_step}
+
+        if speakers is not None:
+            return self._generate_multi_speaker(
+                model, text, speakers, guidance_scale, speed, num_step
+            )
+
+        kwargs = {"text": text, "speed": speed, "num_step": num_step, "guidance_scale": guidance_scale}
 
         if mode == "voice_cloning" and ref_audio is None:
             raise ValueError("voice_cloning mode requires ref_audio to be connected")
@@ -115,17 +158,13 @@ class OmniVoiceGenerate:
             raise ValueError("voice_design mode requires an instruct string (e.g. 'female, low pitch')")
 
         if mode == "voice_cloning":
-            tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
+            tmp_path = _write_tmp_wav(ref_audio)
-            tmp_path = tmp.name
-            tmp.close()
             try:
-                ref_waveform = ref_audio["waveform"].squeeze(0).cpu()  # (channels, samples)
-                audio_np = ref_waveform.numpy()
-                # soundfile expects (samples,) for mono or (samples, channels) for multi-channel
-                sf.write(tmp_path, audio_np[0] if audio_np.shape[0] == 1 else audio_np.T, int(ref_audio["sample_rate"]))
                 kwargs["ref_audio"] = tmp_path
                 if ref_text:
                     kwargs["ref_text"] = ref_text
+                if instruct:
+                    kwargs["instruct"] = instruct
                 audio_tensors = model.generate(**kwargs)
             finally:
                 try:
@@ -140,9 +179,85 @@ class OmniVoiceGenerate:
         else:  # auto_voice or fallback
             audio_tensors = model.generate(**kwargs)
 
-        # Concatenate chunks: each tensor is (1, T) → concat along T → (1, T_total)
+        return self._tensors_to_audio(audio_tensors)
-        combined = torch.cat(audio_tensors, dim=1).cpu()  # (1, T_total) on CPU
-        # ComfyUI AUDIO format: (batch, channels, samples)
-        waveform = combined.unsqueeze(0)  # (1, 1, T_total)
 
+    def _generate_multi_speaker(self, model, text, speakers_data, guidance_scale, speed, num_step):
+        speaker_list = speakers_data["speakers"]
+        spk_mode = speakers_data["mode"]
+        label_map = {s["label"].lower(): i for i, s in enumerate(speaker_list)}
+
+        if spk_mode == "alternate_paragraphs":
+            paragraphs = [p.strip() for p in text.split("\n\n") if p.strip()]
+            if not paragraphs:
+                raise ValueError("OmniVoice Multi-Speaker: no paragraphs found in text.")
+            segments = [
+                (para, speaker_list[i % len(speaker_list)])
+                for i, para in enumerate(paragraphs)
+            ]
+        else:  # tagged_speakers
+            # In tagged mode each line that starts with [Tag] begins a new segment.
+            # Continuation lines (no tag) are appended to the previous segment so
+            # multi-line speeches stay together. Both \n and \n\n separators work.
+            raw_segments: list[list[str]] = []
+            current: list[str] = []
+            for line in text.splitlines():
+                line = line.strip()
+                if not line:
+                    continue
+                if _TAG_RE.match(line):
+                    if current:
+                        raw_segments.append(current)
+                    current = [line]
+                else:
+                    current.append(line)
+            if current:
+                raw_segments.append(current)
+
+            if not raw_segments:
+                raise ValueError("OmniVoice Multi-Speaker: no tagged segments found in text.")
+
+            segments = []
+            for lines in raw_segments:
+                joined = " ".join(lines)
+                m = _TAG_RE.match(joined)
+                if m:
+                    tag = m.group(1).strip().lower()
+                    body = m.group(2).strip()
+                    spk = speaker_list[label_map.get(tag, 0)]
+                else:
+                    body = joined
+                    spk = speaker_list[0]
+                if body:
+                    segments.append((body, spk))
+
+        if not segments:
+            raise ValueError("OmniVoice Multi-Speaker: no text segments to generate.")
+
+        all_chunks = []
+        for para_text, spk in segments:
+            tmp_path = _write_tmp_wav(spk["ref_audio"])
+            try:
+                kwargs = {
+                    "text": para_text,
+                    "ref_audio": tmp_path,
+                    "speed": speed,
+                    "num_step": num_step,
+                    "guidance_scale": guidance_scale,
+                }
+                if spk["ref_text"]:
+                    kwargs["ref_text"] = spk["ref_text"]
+                chunks = model.generate(**kwargs)
+                all_chunks.extend(chunks)
+            finally:
+                try:
+                    os.unlink(tmp_path)
+                except OSError:
+                    pass
+
+        return self._tensors_to_audio(all_chunks)
+
+    @staticmethod
+    def _tensors_to_audio(tensors):
+        combined = torch.cat(tensors, dim=1).cpu()  # (1, T_total)
+        waveform = combined.unsqueeze(0)             # (1, 1, T_total)
         return ({"waveform": waveform, "sample_rate": 24000},)

nodes/loader.py

@@ -1,10 +1,12 @@
 import os
 import torch
 
+_omnivoice_import_error = None
 try:
     from omnivoice import OmniVoice
-except ImportError:
+except Exception as e:
-    OmniVoice = None  # deferred; will raise at runtime if package is missing
+    OmniVoice = None
+    _omnivoice_import_error = e
 
 try:
     import folder_paths
@@ -55,9 +57,14 @@ class OmniVoiceModelLoader:
 
     def load_model(self, device, dtype, compile=False):
         if OmniVoice is None:
-            raise ImportError(
+            msg = (
-                "omnivoice is not installed. Run: pip install omnivoice --no-deps"
+                "omnivoice failed to import. "
+                "Install it with: pip install omnivoice --no-deps\n"
+                "(On Windows embedded Python: .\\python_embeded\\python.exe -m pip install omnivoice --no-deps)\n"
             )
+            if _omnivoice_import_error is not None:
+                msg += f"\nOriginal error: {_omnivoice_import_error}"
+            raise ImportError(msg)
 
         model = OmniVoice.from_pretrained(
             "k2-fsa/OmniVoice",

nodes/multi_speaker.py

@@ -0,0 +1,103 @@
+class OmniVoiceSpeaker:
+    """Bundle a label, reference audio, and optional transcript into a speaker slot."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "label": ("STRING", {
+                    "default": "Narrator",
+                    "tooltip": (
+                        "Name used to identify this speaker.\n"
+                        "In tagged_speakers mode, prefix paragraphs with [Label]:\n"
+                        "  [Narrator] Once upon a time...\n"
+                        "In alternate_paragraphs mode the label is informational only."
+                    ),
+                }),
+                "ref_audio": ("AUDIO", {
+                    "tooltip": "Reference audio clip for this speaker's voice.",
+                }),
+            },
+            "optional": {
+                "ref_text": ("STRING", {
+                    "default": "",
+                    "tooltip": "Transcript of ref_audio. Improves cloning quality.",
+                }),
+            },
+        }
+
+    RETURN_TYPES = ("OMNIVOICE_SPEAKER",)
+    RETURN_NAMES = ("speaker",)
+    FUNCTION = "build"
+    CATEGORY = "OmniVoice"
+
+    def build(self, label, ref_audio, ref_text=""):
+        return ({"label": label, "ref_audio": ref_audio, "ref_text": ref_text},)
+
+
+class OmniVoiceSpeakers:
+    """Collect multiple speakers into a roster for multi-speaker generation.
+
+    The number of speaker input slots expands dynamically when num_speakers changes
+    (requires the OmniVoice web extension to be loaded by ComfyUI).
+    Connect one OmniVoice Speaker node per slot.
+    """
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        # speaker_1…speaker_8 are declared here so ComfyUI validation accepts them.
+        # Visibility is controlled by the JS extension (web/multi_speaker.js):
+        # only the first num_speakers slots are shown as live inputs.
+        optional_speakers = {
+            f"speaker_{i}": ("OMNIVOICE_SPEAKER", {})
+            for i in range(1, 9)
+        }
+        return {
+            "required": {
+                "num_speakers": ("INT", {
+                    "default": 2, "min": 2, "max": 8, "step": 1,
+                    "tooltip": (
+                        "Number of active speaker slots.\n"
+                        "Changing this value adds or removes speaker_N inputs on the node."
+                    ),
+                }),
+                "mode": (
+                    ["alternate_paragraphs", "tagged_speakers"],
+                    {
+                        "default": "alternate_paragraphs",
+                        "tooltip": (
+                            "alternate_paragraphs – paragraphs (separated by blank lines) rotate\n"
+                            "  through speakers in order: 1 → 2 → 3 → 1 → …\n"
+                            "\n"
+                            "tagged_speakers – prefix each paragraph with [Label] to assign\n"
+                            "  a specific speaker. Labels must match those on the Speaker nodes.\n"
+                            "  Unrecognised tags fall back to speaker 1.\n"
+                            "\n"
+                            "  Example:\n"
+                            "  [Narrator] The door creaked open.\n"
+                            "\n"
+                            "  [Alice] Who is there?"
+                        ),
+                    },
+                ),
+            },
+            "optional": optional_speakers,
+        }
+
+    RETURN_TYPES = ("OMNIVOICE_SPEAKERS",)
+    RETURN_NAMES = ("speakers",)
+    FUNCTION = "build"
+    CATEGORY = "OmniVoice"
+
+    def build(self, num_speakers, mode, **kwargs):
+        speakers = []
+        for i in range(1, num_speakers + 1):
+            spk = kwargs.get(f"speaker_{i}")
+            if spk is not None:
+                speakers.append(spk)
+        if len(speakers) < 2:
+            raise ValueError(
+                f"OmniVoice Speakers: at least 2 speakers must be connected "
+                f"(got {len(speakers)})."
+            )
+        return ({"speakers": speakers, "mode": mode},)

nodes/voice_design.py

@@ -0,0 +1,74 @@
+class OmniVoiceVoiceDesign:
+    """Compose a voice design instruct string from structured dropdowns."""
+
+    GENDERS = ["none", "male", "female"]
+
+    AGES = ["none", "child", "teenager", "young adult", "middle-aged", "elderly"]
+
+    PITCHES = [
+        "none",
+        "very low pitch", "low pitch", "moderate pitch",
+        "high pitch", "very high pitch", "whisper",
+    ]
+
+    # Exactly the accents validated by the model's _resolve_instruct() for English
+    ACCENTS = [
+        "none",
+        "american accent", "australian accent", "british accent",
+        "canadian accent", "chinese accent", "indian accent",
+        "japanese accent", "korean accent", "portuguese accent",
+        "russian accent",
+    ]
+
+    # Chinese dialect items validated by the model's _resolve_instruct()
+    ZH_GENDERS  = ["none", "男", "女"]
+    ZH_AGES     = ["none", "儿童", "少年", "青年", "中年", "老年"]
+    ZH_PITCHES  = ["none", "极低音调", "低音调", "中音调", "高音调", "极高音调", "耳语"]
+    ZH_DIALECTS = [
+        "none",
+        "东北话", "云南话", "四川话", "宁夏话", "桂林话",
+        "河南话", "济南话", "甘肃话", "石家庄话", "贵州话",
+        "陕西话", "青岛话",
+    ]
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "language": (
+                    ["English", "Chinese"],
+                    {
+                        "default": "English",
+                        "tooltip": "Selects the instruct vocabulary. The language output wires directly into Generate — no need to set it there too.",
+                    },
+                ),
+                "gender": (cls.GENDERS, {"default": "female",
+                    "tooltip": "Voice gender (English). Ignored when language is Chinese — use zh_gender."}),
+                "age":    (cls.AGES,    {"default": "none",
+                    "tooltip": "Age of the speaker (English). Ignored when language is Chinese — use zh_age."}),
+                "pitch":  (cls.PITCHES, {"default": "none",
+                    "tooltip": "Pitch (English). Ignored when language is Chinese — use zh_pitch."}),
+                "accent": (cls.ACCENTS, {"default": "none",
+                    "tooltip": "Accent (English only, 10 supported values)."}),
+            },
+            "optional": {
+                "zh_gender":  (cls.ZH_GENDERS,  {"default": "none", "tooltip": "声线性别 (Chinese mode)"}),
+                "zh_age":     (cls.ZH_AGES,      {"default": "none", "tooltip": "年龄段 (Chinese mode)"}),
+                "zh_pitch":   (cls.ZH_PITCHES,   {"default": "none", "tooltip": "音调 (Chinese mode)"}),
+                "zh_dialect": (cls.ZH_DIALECTS,  {"default": "none", "tooltip": "方言/口音 (Chinese mode)"}),
+            },
+        }
+
+    RETURN_TYPES  = ("STRING",)
+    RETURN_NAMES  = ("instruct",)
+    FUNCTION      = "compose"
+    CATEGORY      = "OmniVoice"
+
+    def compose(self, language, gender, age, pitch, accent,
+                zh_gender="none", zh_age="none", zh_pitch="none", zh_dialect="none"):
+        if language == "Chinese":
+            parts = [v for v in [zh_gender, zh_age, zh_pitch, zh_dialect] if v != "none"]
+            return ("，".join(parts),)
+        else:
+            parts = [v for v in [gender, age, pitch, accent] if v != "none"]
+            return (", ".join(parts),)

nodes/voice_presets.py

@@ -53,14 +53,45 @@ PRESETS = {
 }
 
 
-def _load_audio(url):
+_AUDIO_EXTS = {".wav", ".flac", ".mp3", ".ogg", ".m4a"}
-    """Download (once) and return (waveform_tensor, sample_rate)."""
+_BUILTIN_FILES = frozenset(os.path.basename(url.split("?")[0]) for url, _ in PRESETS.values())
+
+
+def _scan_user_presets():
+    """Return a dict of user presets found in _CACHE_DIR.
+
+    For each audio file that is not a cached built-in, look for a same-stem
+    .txt file for the transcript.  Key format: "<stem> (local)".
+    """
+    user = {}
+    if not os.path.isdir(_CACHE_DIR):
+        return user
+    for fname in sorted(os.listdir(_CACHE_DIR)):
+        stem, ext = os.path.splitext(fname)
+        if ext.lower() not in _AUDIO_EXTS or fname in _BUILTIN_FILES:
+            continue
+        audio_path = os.path.join(_CACHE_DIR, fname)
+        txt_path = os.path.join(_CACHE_DIR, stem + ".txt")
+        transcript = ""
+        if os.path.exists(txt_path):
+            with open(txt_path, "r", encoding="utf-8") as f:
+                transcript = f.read().strip()
+        user[f"{stem} (local)"] = (audio_path, transcript)
+    return user
+
+
+def _load_audio(source):
+    """Load audio from a URL (downloading once) or a local file path."""
     os.makedirs(_CACHE_DIR, exist_ok=True)
-    filename = os.path.basename(url.split("?")[0])
+    if source.startswith("http://") or source.startswith("https://"):
+        filename = os.path.basename(source.split("?")[0])
         cache_path = os.path.join(_CACHE_DIR, filename)
         if not os.path.exists(cache_path):
-        urllib.request.urlretrieve(url, cache_path)
+            urllib.request.urlretrieve(source, cache_path)
-    audio_np, sr = sf.read(cache_path, dtype="float32")
+        path = cache_path
+    else:
+        path = source
+    audio_np, sr = sf.read(path, dtype="float32")
     if audio_np.ndim == 1:
         audio_np = audio_np[np.newaxis, :]        # (1, samples)
     else:
@@ -72,15 +103,20 @@ def _load_audio(url):
 class OmniVoiceVoicePreset:
     @classmethod
     def INPUT_TYPES(cls):
+        all_presets = {**PRESETS, **_scan_user_presets()}
         return {
             "required": {
                 "preset": (
-                    list(PRESETS.keys()),
+                    list(all_presets.keys()),
                     {
                         "tooltip": (
                             "Pre-fetched reference voice for OmniVoice Generate.\n"
                             "Connect ref_audio → ref_audio and ref_text → ref_text.\n"
-                            "If ref_text is blank, connect a Whisper node to supply the transcript."
+                            "\n"
+                            "To add your own presets, drop audio files into:\n"
+                            f"  {_CACHE_DIR}\n"
+                            "Add a same-name .txt file alongside for the transcript.\n"
+                            "Restart ComfyUI to pick up new files."
                         ),
                     },
                 ),
@@ -93,6 +129,7 @@ class OmniVoiceVoicePreset:
     CATEGORY = "OmniVoice"
 
     def load_preset(self, preset):
-        url, transcript = PRESETS[preset]
+        all_presets = {**PRESETS, **_scan_user_presets()}
-        waveform, sr = _load_audio(url)
+        source, transcript = all_presets[preset]
+        waveform, sr = _load_audio(source)
         return ({"waveform": waveform, "sample_rate": sr}, transcript)

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
-name = "comfyui-omnivoice"
+name = "comfyui-omnivoice-fel"
 description = "ComfyUI nodes for OmniVoice — multilingual zero-shot TTS with voice cloning, voice design, EPUB loading, and voice mixing."
-version = "1.0.0"
+version = "1.0.6"
 license = { text = "GPL-3.0-only" }
 dependencies = []
 

requirements.txt

@@ -1,6 +1,9 @@
-transformers>=4.40.0
+transformers>=5.3.0
 accelerate
-pydub
 soundfile
 numpy
+pydub
+tensorboardx
+webdataset
 beautifulsoup4
+torchcodec

tests/test_epub_loader.py

@@ -58,14 +58,14 @@ def test_input_types_structure():
 
 
 def test_return_types():
-    assert OmniVoiceEpubLoader.RETURN_TYPES == ("STRING", "STRING")
+    assert OmniVoiceEpubLoader.RETURN_TYPES == ("STRING", "STRING", "STRING")
-    assert OmniVoiceEpubLoader.RETURN_NAMES == ("text", "chapter_list")
+    assert OmniVoiceEpubLoader.RETURN_NAMES == ("text", "chapter_title", "chapter_list")
 
 
 def test_chapter_extraction_basic():
     epub = make_fake_epub([("Intro", "<p>Hello world</p>"), ("Chapter One", "<p>Body here</p>")])
     with patch('nodes.epub_loader.zipfile.ZipFile', side_effect=epub_opener(epub)):
-        text, chapter_list = OmniVoiceEpubLoader().load_epub('/fake.epub', 1, 2)
+        text, chapter_title, chapter_list = OmniVoiceEpubLoader().load_epub('/fake.epub', 1, 2)
     assert "Hello world" in text
     assert "Body here" in text
     assert "---" in text
@@ -75,7 +75,7 @@ def test_chapter_extraction_basic():
 def test_chapter_range_single():
     epub = make_fake_epub([("One", "<p>First</p>"), ("Two", "<p>Second</p>"), ("Three", "<p>Third</p>")])
     with patch('nodes.epub_loader.zipfile.ZipFile', side_effect=epub_opener(epub)):
-        text, _ = OmniVoiceEpubLoader().load_epub('/fake.epub', 2, 2)
+        text, _, _ = OmniVoiceEpubLoader().load_epub('/fake.epub', 2, 2)
     assert "Second" in text
     assert "First" not in text
     assert "Third" not in text
@@ -84,7 +84,7 @@ def test_chapter_range_single():
 def test_chapter_list_contains_all():
     epub = make_fake_epub([("A", ""), ("B", ""), ("C", "")])
     with patch('nodes.epub_loader.zipfile.ZipFile', side_effect=epub_opener(epub)):
-        _, chapter_list = OmniVoiceEpubLoader().load_epub('/fake.epub', 2, 2)
+        _, _, chapter_list = OmniVoiceEpubLoader().load_epub('/fake.epub', 2, 2)
     lines = chapter_list.strip().splitlines()
     assert len(lines) == 3
     assert lines[0].startswith("1.")
@@ -94,14 +94,14 @@ def test_chapter_list_contains_all():
 def test_range_clamping_high():
     epub = make_fake_epub([("A", "<p>aaa</p>"), ("B", "<p>bbb</p>")])
     with patch('nodes.epub_loader.zipfile.ZipFile', side_effect=epub_opener(epub)):
-        text, _ = OmniVoiceEpubLoader().load_epub('/fake.epub', 1, 99)
+        text, _, _ = OmniVoiceEpubLoader().load_epub('/fake.epub', 1, 99)
     assert "aaa" in text and "bbb" in text
 
 
 def test_range_clamping_end_below_start():
     epub = make_fake_epub([("A", "<p>aaa</p>"), ("B", "<p>bbb</p>")])
     with patch('nodes.epub_loader.zipfile.ZipFile', side_effect=epub_opener(epub)):
-        text, _ = OmniVoiceEpubLoader().load_epub('/fake.epub', 2, 1)
+        text, _, _ = OmniVoiceEpubLoader().load_epub('/fake.epub', 2, 1)
     assert "bbb" in text
     assert "aaa" not in text
 
@@ -119,14 +119,14 @@ def test_missing_title_fallback():
         z.writestr('OEBPS/ch0.xhtml', '<html><body><p>No title here</p></body></html>')
     buf.seek(0)
     with patch('nodes.epub_loader.zipfile.ZipFile', side_effect=epub_opener(buf.read())):
-        _, chapter_list = OmniVoiceEpubLoader().load_epub('/fake.epub', 1, 1)
+        _, _, chapter_list = OmniVoiceEpubLoader().load_epub('/fake.epub', 1, 1)
     assert "1. Chapter 1" in chapter_list
 
 
 def test_script_style_stripped():
     epub = make_fake_epub([("Test", '<script>alert("xss")</script><style>color:red</style><p>clean</p>')])
     with patch('nodes.epub_loader.zipfile.ZipFile', side_effect=epub_opener(epub)):
-        text, _ = OmniVoiceEpubLoader().load_epub('/fake.epub', 1, 1)
+        text, _, _ = OmniVoiceEpubLoader().load_epub('/fake.epub', 1, 1)
     assert "alert" not in text
     assert "color" not in text
     assert "clean" in text

tests/test_generator.py

@@ -46,7 +46,7 @@ def test_generate_auto_voice():
     assert "sample_rate" in audio
     assert audio["sample_rate"] == 24000
     mock_model.generate.assert_called_once_with(
-        text="Hello world", speed=1.0, num_step=32
+        text="Hello world", speed=1.0, num_step=32, guidance_scale=2.0
     )
 
 
@@ -64,7 +64,7 @@ def test_generate_voice_design():
     audio = result[0]
     assert audio["sample_rate"] == 24000
     mock_model.generate.assert_called_once_with(
-        text="Hello world", instruct="female, low pitch", speed=1.0, num_step=32
+        text="Hello world", instruct="female, low pitch", speed=1.0, num_step=32, guidance_scale=2.0
     )
 
 

web/multi_speaker.js

@@ -0,0 +1,70 @@
+import { app } from "../../scripts/app.js";
+
+const MAX_SPEAKERS = 8;
+
+app.registerExtension({
+    name: "OmniVoice.MultiSpeaker",
+
+    beforeRegisterNodeDef(nodeType, nodeData) {
+        if (nodeData.name !== "OmniVoiceSpeakers") return;
+
+        /**
+         * Ensure the node has exactly `count` speaker_N inputs.
+         * Safe to call multiple times with the same count (idempotent).
+         */
+        function syncSpeakerInputs(node, count) {
+            count = Math.max(2, Math.min(MAX_SPEAKERS, Math.floor(count)));
+
+            // Add any missing slots
+            for (let i = 1; i <= count; i++) {
+                const name = `speaker_${i}`;
+                if (!node.inputs?.find(inp => inp.name === name)) {
+                    node.addInput(name, "OMNIVOICE_SPEAKER");
+                }
+            }
+
+            // Remove excess slots (high → low so indices stay valid)
+            for (let i = MAX_SPEAKERS; i > count; i--) {
+                const name = `speaker_${i}`;
+                const idx = node.inputs?.findIndex(inp => inp.name === name) ?? -1;
+                if (idx === -1) continue;
+                // Sever any connected link before removing the slot
+                const linkId = node.inputs[idx].link;
+                if (linkId != null) node.graph?.removeLink(linkId);
+                node.removeInput(idx);
+            }
+
+            node.setDirtyCanvas(true, true);
+        }
+
+        /**
+         * Attach the num_speakers widget callback once per node instance.
+         * Guarded by a flag so configure() can call it safely on reload.
+         */
+        function attachCallback(node) {
+            if (node._omnivoiceCbAttached) return;
+            const w = node.widgets?.find(w => w.name === "num_speakers");
+            if (!w) return;
+            node._omnivoiceCbAttached = true;
+            w.callback = (value) => syncSpeakerInputs(node, value);
+        }
+
+        // --- Fresh node creation ---
+        const onNodeCreated = nodeType.prototype.onNodeCreated;
+        nodeType.prototype.onNodeCreated = function () {
+            onNodeCreated?.apply(this, arguments);
+            attachCallback(this);
+            const w = this.widgets?.find(w => w.name === "num_speakers");
+            if (w) syncSpeakerInputs(this, w.value);
+        };
+
+        // --- Workflow load: called by LiteGraph after widget values are restored ---
+        const onConfigure = nodeType.prototype.onConfigure;
+        nodeType.prototype.onConfigure = function (data) {
+            onConfigure?.apply(this, arguments);
+            attachCallback(this);
+            const w = this.widgets?.find(w => w.name === "num_speakers");
+            if (w) syncSpeakerInputs(this, w.value);
+        };
+    },
+});

workflows/omnivoice_voice_cloning.json

@@ -1,55 +1,115 @@
 {
   "id": "85925298-bd01-4df7-b8e5-ac92f9d6582e",
   "revision": 0,
-  "last_node_id": 6,
+  "last_node_id": 14,
-  "last_link_id": 9,
+  "last_link_id": 30,
   "nodes": [
     {
-      "id": 1,
+      "id": 6,
-      "type": "OmniVoiceModelLoader",
+      "type": "PreviewAudio",
       "pos": [
-        32,
+        928,
         96
       ],
       "size": [
-        320,
+        448,
-        160
+        96
       ],
       "flags": {},
-      "order": 0,
+      "order": 4,
       "mode": 0,
-      "inputs": [],
+      "inputs": [
-      "outputs": [
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "link": 27
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.18.1",
+        "Node name for S&R": "PreviewAudio",
+        "ue_properties": {
+          "widget_ue_connectable": {},
+          "input_ue_unconnectable": {},
+          "version": "7.8"
+        }
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 14,
+      "type": "OmniVoiceGenerate",
+      "pos": [
+        512,
+        96
+      ],
+      "size": [
+        384,
+        448
+      ],
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [
         {
           "name": "model",
           "type": "OMNIVOICE_MODEL",
-          "slot_index": 0,
+          "link": 23
+        },
+        {
+          "name": "ref_audio",
+          "shape": 7,
+          "type": "AUDIO",
+          "link": 29
+        },
+        {
+          "name": "ref_text",
+          "shape": 7,
+          "type": "STRING",
+          "widget": {
+            "name": "ref_text"
+          },
+          "link": 30
+        }
+      ],
+      "outputs": [
+        {
+          "name": "audio",
+          "type": "AUDIO",
           "links": [
-            5
+            27
           ]
         }
       ],
       "properties": {
-        "aux_id": "ethanfel/ComfyUI-Omnivoice",
+        "cnr_id": "comfyui-omnivoice-fel",
-        "ver": "9b62c9bda8469171025709e2fae3ce877ee0f059",
+        "ver": "340c0aa402fdf8b56fad5eb1559ff901c7fc7cfc",
-        "Node name for S&R": "OmniVoiceModelLoader",
+        "Node name for S&R": "OmniVoiceGenerate",
         "ue_properties": {
           "widget_ue_connectable": {},
-          "version": "7.8",
+          "input_ue_unconnectable": {},
-          "input_ue_unconnectable": {}
+          "version": "7.8"
         }
       },
       "widgets_values": [
-        "cuda:0",
+        "Hello! This is a test of the OmniVoice text-to-speech system.",
-        "float16",
+        "voice_cloning",
-        false
+        "",
+        "",
+        2,
+        1,
+        32,
+        42,
+        "fixed"
       ]
     },
     {
       "id": 2,
       "type": "OmniVoiceVoicePreset",
       "pos": [
-        32,
+        64,
         320
       ],
       "size": [
@@ -57,7 +117,7 @@
         128
       ],
       "flags": {},
-      "order": 1,
+      "order": 0,
       "mode": 0,
       "inputs": [],
       "outputs": [
@@ -66,7 +126,7 @@
           "type": "AUDIO",
           "slot_index": 0,
           "links": [
-            6
+            29
           ]
         },
         {
@@ -74,14 +134,15 @@
           "type": "STRING",
           "slot_index": 1,
           "links": [
-            7
+            30
           ]
         }
       ],
       "properties": {
-        "aux_id": "ethanfel/ComfyUI-Omnivoice",
+        "cnr_id": "comfyui-omnivoice-fel",
         "ver": "9b62c9bda8469171025709e2fae3ce877ee0f059",
         "Node name for S&R": "OmniVoiceVoicePreset",
+        "aux_id": "ethanfel/ComfyUI-Omnivoice",
         "ue_properties": {
           "widget_ue_connectable": {},
           "version": "7.8",
@@ -93,136 +154,124 @@
       ]
     },
     {
-      "id": 5,
+      "id": 1,
-      "type": "OmniVoiceGenerate",
+      "type": "OmniVoiceModelLoader",
       "pos": [
-        416,
+        64,
         96
       ],
       "size": [
-        384,
+        320,
-        448
+        160
+      ],
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "model",
+          "type": "OMNIVOICE_MODEL",
+          "slot_index": 0,
+          "links": [
+            23
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfyui-omnivoice-fel",
+        "ver": "9b62c9bda8469171025709e2fae3ce877ee0f059",
+        "Node name for S&R": "OmniVoiceModelLoader",
+        "aux_id": "ethanfel/ComfyUI-Omnivoice",
+        "ue_properties": {
+          "widget_ue_connectable": {},
+          "version": "7.8",
+          "input_ue_unconnectable": {}
+        }
+      },
+      "widgets_values": [
+        "cuda:0",
+        "float32",
+        false
+      ]
+    },
+    {
+      "id": 13,
+      "type": "OmniVoiceVoiceDesign",
+      "pos": [
+        64,
+        512
+      ],
+      "size": [
+        288,
+        288
       ],
       "flags": {},
       "order": 2,
       "mode": 0,
-      "inputs": [
+      "inputs": [],
-        {
-          "name": "model",
-          "type": "OMNIVOICE_MODEL",
-          "link": 5
-        },
-        {
-          "name": "ref_audio",
-          "shape": 7,
-          "type": "AUDIO",
-          "link": 6
-        },
-        {
-          "name": "ref_text",
-          "shape": 7,
-          "type": "STRING",
-          "widget": {
-            "name": "ref_text"
-          },
-          "link": 7
-        }
-      ],
       "outputs": [
         {
-          "name": "audio",
+          "name": "instruct",
-          "type": "AUDIO",
+          "type": "STRING",
-          "links": [
+          "links": []
-            9
-          ]
         }
       ],
       "properties": {
+        "cnr_id": "comfyui-omnivoice-fel",
+        "ver": "340c0aa402fdf8b56fad5eb1559ff901c7fc7cfc",
+        "Node name for S&R": "OmniVoiceVoiceDesign",
         "ue_properties": {
           "widget_ue_connectable": {},
-          "input_ue_unconnectable": {}
+          "input_ue_unconnectable": {},
-        },
+          "version": "7.8"
-        "aux_id": "ethanfel/ComfyUI-Omnivoice",
+        }
-        "ver": "9b62c9bda8469171025709e2fae3ce877ee0f059",
-        "Node name for S&R": "OmniVoiceGenerate"
       },
       "widgets_values": [
-        "Hello! This is a test of the OmniVoice text-to-speech system.",
+        "English",
-        "voice_cloning",
+        "none",
-        "",
+        "none",
-        "",
+        "none",
-        1,
+        "american accent",
-        32,
+        "none",
-        42,
+        "none",
-        "fixed"
+        "none",
+        "none"
       ]
-    },
-    {
-      "id": 6,
-      "type": "PreviewAudio",
-      "pos": [
-        832,
-        96
-      ],
-      "size": [
-        270,
-        88
-      ],
-      "flags": {},
-      "order": 3,
-      "mode": 0,
-      "inputs": [
-        {
-          "name": "audio",
-          "type": "AUDIO",
-          "link": 9
-        }
-      ],
-      "outputs": [],
-      "properties": {
-        "ue_properties": {
-          "widget_ue_connectable": {},
-          "input_ue_unconnectable": {}
-        },
-        "cnr_id": "comfy-core",
-        "ver": "0.18.1",
-        "Node name for S&R": "PreviewAudio"
-      },
-      "widgets_values": []
     }
   ],
   "links": [
     [
-      5,
+      23,
       1,
       0,
-      5,
+      14,
       0,
       "OMNIVOICE_MODEL"
     ],
     [
+      27,
+      14,
+      0,
       6,
+      0,
+      "AUDIO"
+    ],
+    [
+      29,
       2,
       0,
-      5,
+      14,
       1,
       "AUDIO"
     ],
     [
-      7,
+      30,
       2,
       1,
-      5,
+      14,
       2,
       "STRING"
-    ],
-    [
-      9,
-      5,
-      0,
-      6,
-      0,
-      "AUDIO"
     ]
   ],
   "groups": [],
@@ -231,8 +280,8 @@
     "ds": {
       "scale": 1.201632379715383,
       "offset": [
-        1340.797529638862,
+        1415.539606763376,
-        246.44960712937743
+        123.61214765199416
       ]
     },
     "ue_links": [],