import numpy as np
import soundfile as sf
from faster_qwen3_tts import FasterQwen3TTS

model = FasterQwen3TTS.from_pretrained("lunahr/Celune-1.7B-Upbeat")

buffer = []
for audio_chunk, sr, timing in model.generate_custom_voice_streaming(
    text="Hello, this is Celune.",
    language="Auto",
    speaker="celune"
):
    buffer.append(audio_chunk)

wav = np.concatenate(buffer)
sf.write("output.wav", wav, 24000)

Sample Output

Prompt Steering

You can influence delivery style using text or instruct prompts:

Examples:

“whisper this softly”
“speak with excitement”
“calm and slow delivery”

The model will stay within the defined voice but expand expression.

Notes

This is a single-speaker model. It does not support multiple voices.
Voice identity is fixed and defined by training data.
Avoid mixing languages mid-sentence for best stability.
Avoid special tokens or markup in spoken text.

Intended Use

Character voices
Assistive TTS
Offline voice systems
Creative projects and prototyping

Limitations

Not designed for multi-speaker tasks
Extreme prompts may introduce instability
Long mixed-language inputs may degrade output

About Celune

Celune is a local-first TTS system focused on speed, control, and identity.

Each model in the Celune lineup is:

self-contained
stylistically distinct
built around the same core character concept

License

MIT and Qwen's Apache 2.0 license

Credits

Base model: Qwen3-TTS-12Hz-1.7B-Base
Finetuning + voice design: https://huggingface.co/lunahr