Tutur β€” Lightweight Indonesian Neural TTS (7.62 MB FP16, Pure NumPy CPU)

Tutur adalah model Text-to-Speech (TTS) Bahasa Indonesia ultra-ringan (7.62 MB, di bawah 10 MB) yang berjalan 100% di CPU menggunakan pure NumPy (tanpa GPU, tanpa PyTorch saat inferensi).

Arsitektur & Metode Pelatihan

  • Dataset: X-lord/Dataset-Text-To-Speech-Indonesia (16,38 jam audio studio 22–24 kHz) + referensi Eempostor/F5-TTS-INDO-FINETUNE-V2
  • Teacher Vocoder & Perceptual Feature Matching: Distilasi spektrum kompleks (Log-Magnitude + Instantaneous Phase + Multi-Resolution STFT + Backbone Feature Matching) dari charactr/vocos-mel-24khz (24.000 Hz, 100-bin Mel).
  • Acoustic Model (TuturAcoustic):
    • Phonemic Tokenizer Bahasa Indonesia (membedakan vokal pepet Ι™ vs e serta glottal stop Κ”).
    • ConvNeXt1D Text Encoder + Silence-Anchored Monotonic Alignment Search (MAS).
    • Duration, Pitch ($F_0$), dan Energy Predictor eksplisit.
    • ConvNeXt1D Mel-100 Decoder + 5-layer Harmonic Post-Net.
  • Vocoder (TuturVocos):
    • 8-block ConvNeXt1D Backbone + Exact Inverse STFT (n_fft=1024, hop_length=256, sample_rate=24000).

Contoh Audio

  • sample_1.wav: "Halo, apa kabar kamu hari ini? Aku berharap kamu selalu sehat dan bahagia."
  • sample_2.wav: "Selamat datang di sistem sintesis suara bahasa Indonesia yang ringan dan natural."
  • sample_3.wav: "Kita hidup di zaman perubahan besar di mana kecerdasan buatan berkembang sangat pesat."

Cara Penggunaan (Pure NumPy CPU)

from tutur_engine import IndoNanoSynthesizer as TuturSynthesizer

synth = TuturSynthesizer("tutur_fp16.npz", "config.json")
res = synth.synthesize("Halo selamat pagi Indonesia!")
# res.audio -> float32 numpy array (24,000 Hz)
Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Dataset used to train sansaks/tutur