SobbieTTS

Sobbie is a lightweight, 4.5M-parameter English text-to-speech model built to run fast and completely offline on CPU via ONNX Runtime. It was fine-tuned on an expressive English speech dataset just for fun, starting from the tiny-voca architecture.

Highlights

Parameters ~4.5M (waveform generator included)
Model Size ~21 MB (onnx/model.onnx) Β· ~18 MB (sobbie.pt)
Voices F (primary expressive female) Β· M
Audio Output 24 kHz mono WAV
Hardware Runs fast on CPU (no GPU required)
License Apache-2.0

Download & Setup

Requires Python 3.10+ and git (for installing the offline dizione phonemizer).

# 1. Download the model repository (recommended; handles large weights automatically)
pip install -U huggingface_hub
hf download PatnaikAshish/SobbieTTS --local-dir SobbieTTS
cd SobbieTTS

# (Alternative using git: requires git-lfs)
# git lfs install
# git clone https://huggingface.co/PatnaikAshish/SobbieTTS
# cd SobbieTTS

# 2. Create a virtual environment
python -m venv .venv
# Windows PowerShell:
.\.venv\Scripts\Activate.ps1
# macOS / Linux:
# source .venv/bin/activate

# 3. Install lightweight CPU dependencies
pip install -r requirements.txt

Usage

Command Line (speak.py)

# Speak with voice F (default)
python speak.py --voice F --text "Hello! This is Sobbie running completely offline on your CPU." --output f.wav

# Speak with voice M
python speak.py --voice M --text "Hello! This is Sobbie running completely offline on your CPU." --output m.wav

# Read a multi-sentence text file
python speak.py --voice F --text-file story.txt --output story.wav

Useful Options

Flag Default Description
--voice F Choose F or M
--speed 1.0 Speaking rate (higher = faster)
--pitch-range 1.15 Pitch expressiveness (1.0 = calmer, 1.25 = livelier)
--pitch-shift 0.0 Shift pitch up/down (0.058 β‰ˆ 1 semitone)
--threads 4 CPU threads for ONNX Runtime

Python (ONNX Runtime)

import os
import numpy as np
import onnxruntime as ort
import soundfile as sf

os.environ["DIZIONE_DATA"] = "./phonemizer"
import dizione

phones = dizione.phonemize("Hello! This is Sobbie speaking offline.")
f32 = lambda x: np.array([x], dtype=np.float32)

sess = ort.InferenceSession("onnx/model.onnx", providers=["CPUExecutionProvider"])
audio, _ = sess.run(None, {
    "tokens": np.array([phones.ids], dtype=np.int64),
    "stress": np.array([phones.stress], dtype=np.int64),
    "voice": np.array([0], dtype=np.int64),   # 0 = F, 1 = M
    "speed": f32(1.0),
    "pitch_shift": f32(0.0),
    "pitch_range": f32(1.15),
})

wave = audio[0]
if np.abs(wave).max() > 0.99:
    wave = wave * (0.99 / float(np.abs(wave).max()))
sf.write("hello.wav", wave, 24000)

Run Locally in Your Browser (web/)

Requires Python 3.10+ and git (no virtual environment needed):

cd SobbieTTS\web
python fetch_assets.py
cd public
python -m http.server 8000

Then open http://localhost:8000 in your browser.


Repository Files

  • onnx/model.onnx & onnx/model.json β€” Exported ONNX graph and metadata
  • sobbie.pt β€” PyTorch inference weights
  • phonemizer/ β€” Offline English pronunciation lexicon for dizione
  • web/ β€” Self-contained browser demo (ONNX Runtime Web + Pyodide)
  • speak.py β€” Standalone CPU CLI script
  • config.json & requirements.txt β€” Model configuration and dependencies

Notes & Limitations

  • English only; built as a lightweight experimental/fun fine-tune on an expressive dataset.
  • Unusual foreign names, rare acronyms, and complex spellings depend on the dizione phonemizer.
  • Split very long paragraphs at punctuation marks (handled automatically by speak.py).

Credits & License

  • Built on tiny-voca by Muhammed Nazeem (Apache-2.0) and dizione (Apache-2.0).
  • Released under the Apache-2.0 license.
Downloads last month
75
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for PatnaikAshish/SobbieTTS

Base model

n4ze3m/tiny-voca
Quantized
(3)
this model

Space using PatnaikAshish/SobbieTTS 1