πŸŽ™οΈ Horizon-1M-TTS

An Ultra-Tiny, 1M-Parameter Multi-Speaker Text-to-Speech Model

Params Data Speakers Sample Rate License Status

Built with πŸ–€ by Nebulixlabs

Instagram GitHub HuggingFace


🧠 Model Overview

Horizon-1M-TTS is an ultra-compact, experimental Text-to-Speech (TTS) model trained on 10 hours of speech data. Packing ~1M parameters, it is designed to push the boundaries of extreme model compression β€” making it one of the first ~1M-parameter TTS models capable of generating intelligible, multi-speaker audio across two distinct voices (en_sophia and en_sam).

While the synthesized speech can sound slightly robotic due to its tiny parameter budget and limited training data, Horizon-1M-TTS demonstrates that understandable multi-speaker speech synthesis is possible at a fraction of the size of conventional TTS models β€” ideal for edge-AI research, education, and ultra-low-resource experimentation.


πŸŽ™οΈ Supported Speakers

Speaker ID Language Description
en_sophia English (en) Female voice profile
en_sam English (en) Male voice profile

πŸ—οΈ Model Specifications

Component Value
πŸ”’ Parameters ~1M
πŸ—£οΈ Speakers 2 (en_sophia, en_sam)
🎧 Training Data ~10 hours of English speech
πŸ”Š Sampling Rate 24,000 Hz (default)
🌐 Language English (en)
βš™οΈ Framework PyTorch / Hugging Face transformers (trust_remote_code=True)

πŸ’» How to Use

Prerequisites

pip install torch transformers huggingface_hub soundfile

Inference Code

Note: The snippet below includes a lightweight local patch for a known NameError in modeling_horizon.py before loading the model.

import inspect
import torch
import soundfile as sf
from pathlib import Path
from huggingface_hub import snapshot_download
from transformers import AutoTokenizer, AutoModel

model_id = "Nebulixlabs/Horizon-1M-TTS"
device = "cuda" if torch.cuda.is_available() else "cpu"

# 1. Download the repository locally
local_dir = snapshot_download(repo_id=model_id)

# 2. Patch the NameError in modeling_horizon.py
modeling_file = Path(local_dir) / "modeling_horizon.py"
code = modeling_file.read_text(encoding="utf-8")
buggy_line = "_OrigHorizonForConditionalGeneration = HorizonForConditionalGeneration"
fixed_line = "_OrigHorizonForConditionalGeneration = HorizonTTSModel"

if buggy_line in code:
    modeling_file.write_text(code.replace(buggy_line, fixed_line), encoding="utf-8")

# 3. Load the model
model = AutoModel.from_pretrained(local_dir, trust_remote_code=True).to(device)
model.eval()

# 4. Choose speaker: "en_sophia" or "en_sam"
speaker = "en_sophia"  # or "en_sam"
text = "Hello, this is a test of the Horizon 1M TTS model."

# 5. Run inference (direct text method or tokenizer fallback)
with torch.no_grad():
    if hasattr(model, "synthesize"):
        audio = model.synthesize(text, speaker=speaker)
    elif hasattr(model, "generate"):
        sig = inspect.signature(model.generate)
        if "text" in sig.parameters:
            audio = model.generate(text=text, speaker=speaker)
        else:
            # Fallback to slow tokenizer
            tokenizer = AutoTokenizer.from_pretrained(
                local_dir, trust_remote_code=True, use_fast=False
            )
            inputs = tokenizer(text, return_tensors="pt").to(device)
            if "speaker" in sig.parameters:
                audio = model.generate(**inputs, speaker=speaker)
            else:
                if hasattr(model, "_switch_profile"):
                    model._switch_profile(speaker)
                audio = model.generate(**inputs)
    else:
        raise RuntimeError(
            f"No inference method found. Available methods: {[m for m in dir(model) if not m.startswith('_')]}"
        )

# 6. Save the generated audio
audio_np = audio.cpu().numpy().squeeze() if isinstance(audio, torch.Tensor) else audio
sampling_rate = getattr(model.config, "sampling_rate", 24000)
output_file = f"output_{speaker}.wav"
sf.write(output_file, audio_np, sampling_rate)

print(f"Done! Audio saved to '{output_file}'.")

Usage tips:

  • Keep input sentences clear and concise for the most intelligible pronunciation.
  • Runs comfortably on both CPU and GPU thanks to its ~1M parameter footprint.

🎯 Intended Use

  • βœ… Educational purposes β€” exploring ultra-small neural TTS architectures and multi-speaker conditioning.
  • βœ… Edge-AI and low-resource prototyping where memory and compute are strictly constrained.
  • βœ… Research baseline for extreme parameter-efficiency in speech synthesis.

🚫 Out-of-Scope Use

  • ❌ Production voice assistants, commercial audiobooks, or studio-grade voiceovers.
  • ❌ High-stakes, medical, legal, or safety-critical announcements where crystal-clear natural prosody is mandatory.
  • ❌ Voice cloning or impersonation of real individuals.

⚠️ Limitations & Biases

  • Robotic artifacts: With only ~1M parameters and 10 hours of training data, the generated audio can sound noticeably robotic or metallic compared to larger TTS models.
  • English only: Trained exclusively on English speech; non-English words, rare proper nouns, or complex abbreviations may be mispronounced.
  • Limited expressiveness: Emotional range, natural pauses, and complex prosody are limited.
  • Experimental status: Requires trust_remote_code=True and the temporary modeling_horizon.py patch shown above.

πŸ“œ Citation

@misc{horizon1mtts2026,
  title  = {Horizon-1M-TTS: An Ultra-Tiny Multi-Speaker Text-to-Speech Model},
  author = {Nebulixlabs},
  year   = {2026},
  howpublished = {\url{[https://huggingface.co/Nebulixlabs/Horizon-1M-TTS](https://huggingface.co/Nebulixlabs/Horizon-1M-TTS)}}
}

πŸ”— Connect with Nebulixlabs

Instagram GitHub

Made with πŸ–€ by Nebulixlabs
Downloads last month
-
Safetensors
Model size
2M params
Tensor type
F32
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support