2e3c357e5a
Speed values below 0.3 produce noise, and 0.1 generates 10x the normal audio length which can consume 20+ GB VRAM and freeze the system. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
264 lines
11 KiB
Python
264 lines
11 KiB
Python
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
|
||
def INPUT_TYPES(cls):
|
||
return {
|
||
"required": {
|
||
"model": ("OMNIVOICE_MODEL", {
|
||
"tooltip": "OmniVoice model loaded by the OmniVoice Model Loader node.",
|
||
}),
|
||
"text": ("STRING", {
|
||
"multiline": True,
|
||
"default": "",
|
||
"tooltip": (
|
||
"Text to synthesize. Supports inline tags for expression and pronunciation:\n"
|
||
"\n"
|
||
"NON-VERBAL SOUNDS:\n"
|
||
" [laughter] – insert a laugh\n"
|
||
" [sigh] – insert a sigh\n"
|
||
"\n"
|
||
"QUESTION / CONFIRMATION:\n"
|
||
" [question-en] – rising English question intonation\n"
|
||
" [confirmation-en] – confirmation sound\n"
|
||
"\n"
|
||
"SURPRISE:\n"
|
||
" [surprise-ah] [surprise-oh] [surprise-wa] [surprise-yo]\n"
|
||
"\n"
|
||
"DISSATISFACTION:\n"
|
||
" [dissatisfaction-hnn]\n"
|
||
"\n"
|
||
"ENGLISH PRONUNCIATION (CMU phoneme override):\n"
|
||
" You could probably still make [IH1 T] look good.\n"
|
||
"\n"
|
||
"CHINESE PRONUNCIATION (pinyin + tone number):\n"
|
||
" 严重SHE2本了\n"
|
||
"\n"
|
||
"EXAMPLE:\n"
|
||
" [laughter] You really got me. I didn't see that coming at all."
|
||
),
|
||
}),
|
||
"mode": (
|
||
["voice_cloning", "voice_design", "auto_voice"],
|
||
{
|
||
"default": "voice_cloning",
|
||
"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\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.",
|
||
}),
|
||
"ref_text": ("STRING", {
|
||
"default": "",
|
||
"tooltip": "Transcription of ref_audio. Connect a Whisper (or other STT) node for best results.",
|
||
}),
|
||
"instruct": ("STRING", {
|
||
"default": "",
|
||
"tooltip": (
|
||
"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 pitch, low pitch, moderate pitch, high pitch, very high pitch, whisper\n"
|
||
"\n"
|
||
"ACCENTS (only these are supported by the model):\n"
|
||
" american accent, australian accent, british accent, canadian accent,\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": 0.3, "max": 3.0, "step": 0.1,
|
||
"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,
|
||
"tooltip": "Diffusion steps. 32 = default quality. 16 = faster, slightly lower quality.",
|
||
}),
|
||
"seed": ("INT", {
|
||
"default": 0, "min": 0, "max": 2**32 - 1,
|
||
"tooltip": (
|
||
"Random seed for the diffusion sampler. "
|
||
"Set the same value across all Generate nodes in an audiobook pipeline "
|
||
"to keep the voice consistent between paragraphs/chapters. "
|
||
"0 = random (different each run)."
|
||
),
|
||
}),
|
||
},
|
||
}
|
||
|
||
RETURN_TYPES = ("AUDIO",)
|
||
RETURN_NAMES = ("audio",)
|
||
FUNCTION = "generate"
|
||
CATEGORY = "OmniVoice"
|
||
|
||
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)
|
||
|
||
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")
|
||
if mode == "voice_design" and not instruct:
|
||
raise ValueError("voice_design mode requires an instruct string (e.g. 'female, low pitch')")
|
||
|
||
if mode == "voice_cloning":
|
||
tmp_path = _write_tmp_wav(ref_audio)
|
||
try:
|
||
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:
|
||
os.unlink(tmp_path)
|
||
except OSError:
|
||
pass
|
||
|
||
elif mode == "voice_design" and instruct:
|
||
kwargs["instruct"] = instruct
|
||
audio_tensors = model.generate(**kwargs)
|
||
|
||
else: # auto_voice or fallback
|
||
audio_tensors = model.generate(**kwargs)
|
||
|
||
return self._tensors_to_audio(audio_tensors)
|
||
|
||
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},)
|