This is a fine-tuned CSM (Conversational Speech Model) for Indonesian Text-to-Speech (TTS) generation. The model has been adapted from the original CSM-1B to generate natural-sounding Indonesian speech from text input.

Key Features:

Multi-speaker support (81 unique speakers from the training dataset)
High-quality 24kHz audio output
Fast inference with optimized architecture
Support for various Indonesian text inputs

This CSM model was trained 2x faster with Unsloth and Huggingface's TRL library.

Usage

Installation

pip install transformers torch soundfile ipython

Basic Usage

python

from transformers import CsmForConditionalGeneration, AutoProcessor
import torch
from IPython.display import Audio, display
import soundfile as sf

# Load model dan processor
model = CsmForConditionalGeneration.from_pretrained("Ellbendls/csm-1b-indonesian-fine-tuned")
processor = AutoProcessor.from_pretrained("Ellbendls/csm-1b-indonesian-fine-tuned")

# Move to GPU if available
device = "cuda" if torch.cuda.is_available() else "cpu"
model = model.to(device)

# Generate audio
text = "Selamat pagi, nama saya adalah Budi. Bagaimana kabar Anda hari ini?"
speaker_id = 0  # Use speaker ID 0-80

inputs = processor(f"[{speaker_id}]{text}", add_special_tokens=True).to(device)
audio_values = model.generate(
    **inputs,
    max_new_tokens=125,  # ~10 seconds of audio
    output_audio=True
)

# Convert and save audio
audio = audio_values[0].to(torch.float32).cpu().numpy()
sf.write("indonesian_tts_output.wav", audio, 24000)
display(Audio(audio, rate=24000))

Advanced Usage with Quality Control

python

# Generate with custom parameters for better quality
audio_values = model.generate(
    **inputs,
    max_new_tokens=200,  # For longer text
    depth_decoder_top_p=0.9,
    depth_decoder_do_sample=True,
    depth_decoder_temperature=0.9,
    output_audio=True
)

Model Details

Training Configuration

Base Model: unsloth/csm-1b
Training Steps: 100
Batch Size: 1 (with gradient accumulation steps: 8)
Learning Rate: 1e-4
Optimizer: AdamW 8-bit
Scheduler: Cosine
LoRA Configuration: r=32, alpha=32

Dataset Information

Dataset: octava/indonesian-voice-transcription-1.1.85
Language: Indonesian
Speakers: 81 unique speakers
Audio Duration: 0.55 - 13.8 seconds per sample
Text Length: 5 - 147 characters per sample
Sample Rate: 24kHz

Parameters

Speaker IDs: Use integer values from 0-80 to select different voice characteristics
Max Tokens: 125 tokens ≈ 10 seconds of audio (adjust for longer speech)
Audio Format: 24kHz WAV output
Device: Supports both CPU and GPU inference

Limitations

Optimized for Indonesian language text
Speaker consistency depends on the selected speaker_id
Audio quality may vary with very long text inputs
Best performance with modern Indonesian text (not archaic forms)

Citation

bibtex

@model{csm-1b-indonesian-fine-tuned,
  title={CSM-1B Indonesian Fine-tuned TTS Model},
  author={Ellbendls},
  year={2025},
  base_model={unsloth/csm-1b},
  dataset={octava/indonesian-voice-transcription-1.1.85}
}

Acknowledgments

Thanks to Unsloth for the efficient training framework
Original CSM-1B model developers
octava for providing the Indonesian voice transcription dataset