PiVo-1.4M

PiVo-1.4M

Speaks English in two synthetic voices, Pi and Vo, from a complete text-to-speech model of about 1.4 million parameters that runs on a CPU with no network connection. A research model.

Made by empero-ai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: PiVo-1.4M is a tiny text-to-speech model that reads English aloud in two voices, Pi and Vo. The whole stack, vocoder included, has 1,401,562 learned parameters and writes 16 kHz mono audio.

Speech is generated on your own CPU, using NumPy and ONNX Runtime. PyTorch, a teacher model, a speaker encoder and a network connection are all unnecessary when you generate speech.

Who it suits: Developers and researchers who want speech output from a tiny model on ordinary hardware, and people studying how few parameters a voice needs.

Industry | Who uses it | How they use it | Example
Software | Developers of offline apps | Add local speech output without a GPU | Read a notification aloud on a CPU
Research | Speech synthesis researchers | Study a tiny acoustic model and vocoder | Compare its clarity with larger systems
Hardware | Teams with limited compute | Run speech on small machines | Use the ONNX files in a lightweight program

Skip it if you need languages other than English or the most natural sound, since the card calls it a research model with a somewhat robotic texture on some sentences.

Where to use it:

• A CPU on Windows, macOS or Linux with Python 3.11 or newer
• A terminal, through the inference script
• Python programs, through the PiVo class
• A companion code repository, which the card says adds a command-line tool, an HTTP server and a Dockerfile

Use case | Who | What they get back
Offline narration | Developers | A 16 kHz mono wav file from your text
Streaming playback | App builders | Audio in 10 ms frames, so playback starts before the sentence ends
Custom pronunciation | Anyone with unusual names | Spoken output using your ARPAbet spelling

How to set it up:

• Python 3.11 or newer
• A clone of the model repository, since the commands run from its folder
• The packages in its requirements file

1. Create a virtual environment.
2. Activate it.
3. Install the requirements.
4. Run the synthesize command for the voice you want.

python -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
python inference.py synthesize --model-dir . --voice Pi \
  --text "NASA is ready for the next mission." --output pi.wav
python inference.py synthesize --model-dir . --voice Vo \
  --text "Hello, my name is Vo." --output vo.wav

How to use it: Run the synthesize command with a voice, a text and an output file, or call the PiVo class from Python.

from pivo import PiVo

model = PiVo(".", threads=1)
model.save("pi.wav", "NASA is ready for the next mission.", voice="Pi")

audio = model.synthesize("Hello, my name is Vo.", voice="Vo", speed=1.0)
# float32 mono PCM at model.sample_rate (16000 Hz)

Tool | What it does
inference.py synthesize | Command line route that writes a wav file for a chosen voice
PiVo class | Python entry point, started with the model folder and a thread count
model.save | Writes a wav file from text and a voice
model.synthesize | Returns audio samples, with a speed setting from 0.5 to 2.0
iter_audio | Yields audio in 10 ms frames while the rest is generated

Example: You run the first synthesize command with the voice Pi and the sentence about NASA being ready for the next mission. It writes pi.wav, a 16 kHz mono recording of that sentence in the Pi voice.

For newcomers and experts: If you are new to speech models, start with the two command line examples and listen to the wav files. Keep sentences short, because one utterance can run at most 30 seconds.

If you know the field, an acoustic model of 584,965 parameters predicts vocoder features, and a vocoder of 816,597 parameters derived from FARGAN turns them into audio. The frontend uses the CMU pronouncing dictionary and raises an error for words it does not know.

Text is cleaned and looked up in a dictionary
Acoustic model predicts vocoder features
FARGAN-derived vocoder makes audio
16 kHz mono wav file

Good to know:

• Pronunciation overrides use ARPAbet, so unknown words need your own spelling of the sounds
• Acronyms like CPU are read letter by letter, but NASA is pronounced as one word
• A single request may hold up to 2,000 characters or 512 frontend tokens unless you change the defaults, so break longer text into sentences
• The word error rate shows the words come through clearly, but it does not measure naturalness
• Both voices are synthetic, designed and recorded with Qwen3-TTS models

Title: Word error rate on held-out sentences, lower is better
Validation: 0.863 percent
Test: 0.532 percent

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the text you give it and the files you download, with no network connection needed to speak
• Not for: languages other than English, or loading through the Transformers AutoModel class or a hosted inference service

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: onnx
• Downloads in the last 30 days: 0
• Likes: 59
• License: apache-2.0
• Languages: en
• Quantized versions: 1
• Tags: onnx, pivo, fargan, english, text-to-speech, small-model

Links

• Model: https://huggingface.co/empero-ai/PiVo-1.4M
• Files: https://huggingface.co/empero-ai/PiVo-1.4M/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download empero-ai/PiVo-1.4M

Then follow the usage in its model card. Nothing in this kit runs by itself.