Qwen3-TTS-12Hz-1.7B-Sonia is a custom Sonia voice text-to-speech model derived from a Qwen3 TTS 12Hz 1.7B-class base model. It is intended for local-first voice synthesis inside the SONIA assistant stack and related experimental speech-generation workflows.

This repository contains the Sonia-adapted TTS checkpoint and supporting files for inference, evaluation, and integration.

Status: experimental / research-grade
Primary use: Sonia voice synthesis
Model type: text-to-speech / speech synthesis
Format: Safetensors
Language: English

Model Details

Model name: Qwen3-TTS-12Hz-1.7B-Sonia
Architecture family: Qwen3 TTS
Approximate parameter class: 1.7B
Audio rate / acoustic frame setting: 12Hz TTS configuration
Task: text-to-speech generation
Primary persona: Sonia custom assistant voice
Repository owner: VocaborSilentii
Intended runtime: local inference, SONIA voice service, experimental TTS pipelines

This model is intended to provide SONIA with a consistent synthetic voice for assistant responses, narration, and local voice interaction.

Intended Use

This model is designed for:

Local SONIA voice output
Assistant speech synthesis
Prototype TTS workflows
Offline or local-first AI assistant experiments
Controlled narration and spoken-response generation
Research into custom assistant voice identity

Recommended use cases include short-form assistant replies, guided narration, interface feedback, and generated speech for personal AI workflows.

Out-of-Scope Use

Do not use this model for:

Impersonating real people without consent
Deceptive synthetic media
Fraud, scams, or social engineering
Political persuasion using undisclosed synthetic speech
Harassment, abuse, or non-consensual voice generation
Generating speech that falsely claims to be a real person

This model should be used transparently as synthetic speech.

Training / Fine-Tuning Notes

This model is a Sonia-specific TTS adaptation. It was created to give the SONIA assistant a recognizable, consistent voice profile.

Training details should be completed as available:

Base model: Qwen/Qwen3-TTS-12Hz-1.7B
Fine-tuning method: custom Sonia voice adaptation
Checkpoint source: Sonia TTS fine-tune output
Training data: private / controlled Sonia voice dataset
Dataset release: not included in this repository unless explicitly added
Evaluation status: pending or manually validated

If releasing this model publicly, only include training data details that are legally and ethically safe to disclose.

Recommended Inference Pattern

A typical Sonia voice pipeline should use:

text

Input text
  ↓
Text normalization
  ↓
Qwen3-TTS-12Hz-1.7B-Sonia
  ↓
Audio token / acoustic generation
  ↓
Vocoder or decoder stage
  ↓
Final waveform output

For best quality, normalize input before synthesis:

Expand abbreviations
Remove unsupported symbols
Keep sentences reasonably short
Avoid malformed punctuation
Split long responses into smaller speech segments

Example input:

Sonia is online. Local voice synthesis is active.
Example Usage

The exact inference code depends on the runtime used in your SONIA stack.

Example conceptual usage:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "VocaborSilentii/Qwen3-TTS-12Hz-1.7B-Sonia"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    device_map="auto"
)

text = "Sonia is online. Voice synthesis is ready."

inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.7,
        top_p=0.9
    )

Depending on the final Qwen3 TTS implementation, additional tokenizer, acoustic codec, vocoder, or decoder components may be required.

SONIA Integration

Recommended SONIA provider name:

tts:
  provider: sonia3_tts
  model_id: VocaborSilentii/Qwen3-TTS-12Hz-1.7B-Sonia
  model_type: qwen3_tts
  sample_mode: local
  voice_name: Sonia
  trust_remote_code: true

Recommended runtime role:

sonia_voice:
  role: assistant_tts
  model: VocaborSilentii/Qwen3-TTS-12Hz-1.7B-Sonia
  use_for:
    - assistant_replies
    - narration
    - voice_interface
    - local_speech_output
  safety:
    synthetic_voice_disclosure: true
    impersonation_protection: true
Limitations

This model may have limitations including:

Possible pronunciation errors
Unstable prosody on very long text
Inconsistent emotional tone
Sensitivity to punctuation and formatting
Possible hallucinated acoustic artifacts
Need for additional decoder/vocoder components
Limited multilingual support unless explicitly fine-tuned

For production use, evaluate the model across pronunciation, clarity, latency, stability, and safety behavior.

Safety and Ethics

This model produces synthetic speech. Users should disclose that generated audio is AI-generated when appropriate.

Responsible use requirements:

Do not clone or imitate private individuals without permission.
Do not use the model to deceive listeners.
Do not use generated speech for fraud, harassment, or manipulation.
Do not present Sonia-generated audio as a real human recording.
Respect all applicable laws and platform policies.
Evaluation

Suggested evaluation categories:

Category	Description
Intelligibility	Are words understandable?
Naturalness	Does the voice sound smooth and coherent?
Speaker consistency	Does the Sonia voice remain stable?
Prosody	Are rhythm, pauses, and emphasis appropriate?
Long-text stability	Does quality degrade over longer passages?
Latency	Is generation fast enough for assistant use?
Safety	Does the system avoid misuse scenarios?

Formal benchmark results are not yet included.

Citation

If you use this model, cite this repository:

@misc{qwen3_tts_12hz_1_7b_sonia,
  title        = {Qwen3-TTS-12Hz-1.7B-Sonia},
  author       = {VocaborSilentii},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/VocaborSilentii/Qwen3-TTS-12Hz-1.7B-Sonia}}
}
License

License terms depend on the base model license, training data permissions, and any additional Sonia-specific modifications.

Before public or commercial use, verify:

The base model license
Dataset rights
Voice-data consent
Redistribution permissions
Commercial-use permissions

If unsure, treat this model as research / personal-use only until licensing is confirmed.

Disclaimer

This model is experimental and provided without warranty. Generated speech may contain errors, artifacts, or unintended outputs. Users are responsible for validating outputs and complying with applicable laws, licenses, and ethical requirements.


After pasting it, I recommend setting the visible metadata fields like this:

```text id="hf-fields"
License: Other, unless you have confirmed the exact base-model license
Language: English
Base model: Qwen/Qwen3-TTS-12Hz-1.7B, only if that exact repo exists and is correct
Pipeline tag: Text-to-Speech
Library: Transformers
Tags: qwen3, tts, sonia, custom-voice, speech-synthesis, safetensors