BarryFutureman/NeuTTS-Express

NeuTTS Express

• Compact model size (0.5B)
• Emotional and paralinguistic controls via (tags)
• Supports streaming for low latency usecases
• Instant voice cloning

Supported tags: (neutral), (happy), (sad), (angry), (surprised), (disgusted), (fearful), (giggle), (gasp), (exhale), (laugh), (chuckle), (sign)

import time
import re
import types
import soundfile as sf
from neuttsair.neutts import NeuTTSAir

# patch function
def patched_to_phones(self, text: str) -> str:
    pattern = r'\([^)]*\)'
    preserved_matches = list(re.finditer(pattern, text))
    
    if not preserved_matches:
        phones = self.phonemizer.phonemize([text])[0]
        return " ".join(phones.split())
    
    parts = re.split(pattern, text)
    preserved_texts = [m.group() for m in preserved_matches]

    phonemized_parts = []
    for part in parts:
        if part.strip():
            phonemized = self.phonemizer.phonemize([part])[0]
            phonemized = " ".join(phonemized.split())
            phonemized_parts.append(phonemized)
        else:
            phonemized_parts.append("")

    result = []
    for i, phonemized_part in enumerate(phonemized_parts):
        if phonemized_part:
            result.append(phonemized_part)
        if i < len(preserved_texts):
            result.append(preserved_texts[i])
    
    return " ".join(result)

# apply monkey patch
NeuTTSAir._to_phones = patched_to_phones

# inference
tts = NeuTTSAir(
    backbone_repo="BarryFutureman/NeuTTS-Express",
    backbone_device="cuda",
    codec_repo="neuphonic/distill-neucodec",
    codec_device="cuda"
)

input_text = "I just got the best news ever (giggle), and I couldn't be happier!"
start_time = time.time()
wav = tts.infer(input_text)
end_time = time.time()

sf.write("output.wav", wav, 24000)

Higher temperature, top-p, top-k is recommended.