Blame for README.md

---
license: apache-2.0
language:
  - en
library_name: onnx
pipeline_tag: text-to-speech
tags:
  - text-to-speech
  - tts
  - speech-synthesis
  - neural-tts
  - offline-tts
  - on-device
  - edge-ai
  - edge
  - offline
  - cpu
  - onnx
  - onnxruntime
  - lightweight
  - small-model
  - flow-matching
  - vocos
  - english
  - raspberry-pi
  - flockspeech
  - accessibility
  - screen-reader
  - voice-assistant
  - narration
datasets:
  - keithito/lj_speech
---

# Junco v0.1: tiny offline text-to-speech (TTS) for CPU

**A small, natural English text-to-speech model that runs offline on any CPU. 44 MB, no GPU, no cloud.**

Junco is an 11.4M-parameter flow-matching TTS model from [Flock Speech](https://huggingface.co/flockspeech). It turns
English text into natural 24 kHz speech on an ordinary CPU, faster than real time. It has no GPU, network or PyTorch
dependency, and the full model is a 44 MB download.

Junco is the clear, neutral **reading voice** of the Flock Speech family: a good fit for voice assistants, screen
readers, notifications, narration and other on-device or edge AI speech output. Install it with
`pip install flockspeech`.

| Flock Speech voices | Style | Model card |
|---|---|---|
| **Junco** | clear, neutral reading voice | this page |
| **Goldfinch** | conversational voice that laughs and breathes on cue with `[laugh]` tags | [flockspeech/goldfinch](https://huggingface.co/flockspeech/goldfinch) |

| | |
|---|---|
| **Developed by** | Flock Speech |
| **Model type** | Non-autoregressive TTS: transformer encoder, duration predictor, conditional flow-matching decoder, Vocos vocoder |
| **Language** | English (US) |
| **Voices** | 1 (female, read-aloud style) |
| **Output** | 24 kHz mono audio |
| **Parameters** | 11.4M acoustic model + 13.5M vocoder |
| **Download** | 44 MB (ONNX; int8 decoder, fp16 encoder and vocoder) |
| **Runtime** | ONNX Runtime on CPU (Linux, macOS, Windows; x86-64 and ARM64) |
| **License** | Apache-2.0 (vocoder weights MIT) |
| **Version** | v0.1, November 2026 |

## Samples

All samples were made with the files in this repository at default settings (8 steps, seed 0), with no editing.

| Use | Text | Audio |
|---|---|---|
| Greeting | *Hello! I'm Junco, a tiny voice that runs entirely on your own device.* | <audio controls src="https://huggingface.co/flockspeech/junco/resolve/main/samples/greeting.wav"></audio> |
| Assistant | *Your next meeting starts in fifteen minutes. Would you like me to read the agenda?* | <audio controls src="https://huggingface.co/flockspeech/junco/resolve/main/samples/assistant.wav"></audio> |
| Narration | *The old lighthouse keeper climbed the spiral stairs one last time, and watched the ships come home through the fog.* | <audio controls src="https://huggingface.co/flockspeech/junco/resolve/main/samples/story.wav"></audio> |
| News | *Researchers announced on Tuesday that the new battery can be charged in under ten minutes, and lasts twice as long.* | <audio controls src="https://huggingface.co/flockspeech/junco/resolve/main/samples/news.wav"></audio> |

## Quick start

The [`flockspeech`](https://pypi.org/project/flockspeech/) Python package bundles these weights, so nothing is downloaded
at run time.

```sh
pip install flockspeech
flockspeech "Hello! I'm Junco." -o hello.wav --play
```

```python
from flockspeech import Junco

tts = Junco()                        # loads in under a second
audio = tts("Hello there.")          # float32 numpy array, 24 kHz mono
tts.save("Hello there.", "hello.wav")

# Low latency: audio is produced sentence by sentence
for chunk in tts.stream(long_text):
    play(chunk)
```

To use the files in this repository directly:

```python
from huggingface_hub import snapshot_download
from flockspeech import Junco

path = snapshot_download("flockspeech/junco", allow_patterns=["onnx/*"])
tts = Junco(model_dir=f"{path}/onnx")
```

### Options

| Parameter | Default | Effect |
|---|---|---|
| `speed` | 1.0 | Speaking rate; 1.2 is 20% faster |
| `steps` | 8 | Flow-matching steps; 4 is about 1.5× faster with slightly lower quality |
| `seed` | 0 | Fixes the small random variation in delivery, so output is reproducible |
| `threads` | up to 4 | CPU threads for ONNX Runtime |

## Intended use

**Suitable for:**
- Voice output on devices and in apps that must work offline or keep text private: assistants, accessibility and
  screen-reading tools, kiosks, robotics, games and embedded products.
- CPU-only servers where cost or latency matters.
- Prototyping and research on small TTS models.

**Out of scope:**
- Languages other than English, or speaker voices other than the one provided. Junco has no voice cloning.
- Expressive, emotional or conversational styles. The voice reads in a neutral, read-aloud style.
- Impersonating a real person, or producing audio meant to deceive listeners about who is speaking or whether the
  speech is synthetic.

## Limitations

- **No text normalisation.** Write numbers, currency, dates and abbreviations in words: "five dollars", not "$5";
  "Doctor", not "Dr.". Raw digits and symbols are read by espeak-ng's rules, which may not be what you want.
- **Pronunciation comes from espeak-ng.** Rare names, brands and some homographs, such as *read* and *lead*, can be
  mispronounced.
- **One voice, one accent.** The model learned from a single US English speaker.
- **Long-form audio.** Text is synthesised sentence by sentence. Prosody does not carry across sentences.
- **Bright high frequencies.** About 14% of Junco's energy lies above 4 kHz, against 2% for real recordings. On some
  playback systems this may sound slightly bright or hissy.

## Risks and responsible use

Synthetic speech can be misused for deception. Junco produces a single fixed voice and cannot clone speakers, which
limits impersonation risk. We still ask that you:
- tell listeners when speech is synthetic, where that matters;
- not use the model for fraud, harassment or deceptive impersonation;
- review output before you use it in safety-critical or high-stakes settings.

## Model details

### Architecture

The model runs in four stages:
1. **Text front end:** espeak-ng converts text to IPA phonemes (en-us, with stress marks). A blank token is inserted
   between phonemes.
2. **Text encoder (3.2M parameters):** 3 ConvNeXt pre-net blocks, then 6 transformer layers (width 192, RoPE, SwiGLU,
   RMSNorm). It predicts a per-phoneme mel prior.
3. **Duration predictor (0.3M parameters):** a convolutional regressor of log-durations. It was trained on alignments from Monotonic
   Alignment Search, so no external aligner is needed.
4. **Decoder (7.9M parameters):** a flow-matching U-Net hybrid.
   - ConvNeXt blocks run at the full frame rate, and four adaLN transformer blocks run at half the frame rate.
   - It uses classifier-free guidance (weight 1.0) and 8 Euler steps with a sway-sampling schedule.
   - It outputs a 100-band log-mel spectrogram (hop 256, 24 kHz).

The vocoder is [Vocos](https://huggingface.co/charactr/vocos-mel-24khz) (charactr/vocos-mel-24khz, MIT). In ONNX it
predicts the STFT magnitude and phase, and the inverse STFT runs in numpy.

### Files

| File | Contents | Size |
|---|---|---:|
| `onnx/encoder.onnx` | Text encoder and duration predictor (fp16 weights, fp32 compute) | 7.2 MB |
| `onnx/decoder-int8.onnx` | One guided flow-matching step (int8 dynamic quantisation) | 9.9 MB |
| `onnx/vocoder.onnx` | Vocos backbone and spectral head (fp16 weights, fp32 compute) | 27.1 MB |
| `onnx/junco.json` | Mel normalisation statistics and default inference settings | <1 kB |
| `samples/` | The audio samples above | |

### Training

| | |
|---|---|
| **Data** | [LJ Speech](https://keithito.com/LJ-Speech-Dataset/): 13,100 clips, about 24 hours, one speaker, public domain |
| **Objective** | Conditional flow matching (logit-normal time sampling, σ_min 1e-4), MAS prior loss, duration MSE |
| **Guidance** | Condition dropped for 10% of training examples, for classifier-free guidance |
| **Optimiser** | AdamW (lr 2e-4, β 0.9/0.98, weight decay 0.01), 1,000 warm-up steps, EMA 0.999 |
| **Batching** | Dynamic batches of up to 20,000 mel frames |
| **Compute** | 44,485 steps in 3 hours on one NVIDIA T4 GPU |
| **Released weights** | EMA weights |

**Environmental impact.** The final training run used about 3 GPU-hours on one NVIDIA T4 (70 W TDP), or about
0.2 kWh. Architecture experiments before it used more.

## FAQ

**Does Junco need a GPU or an internet connection?**
No. It runs on an ordinary CPU with ONNX Runtime. Once installed, speech is generated fully offline and your text
never leaves the machine.

**How big is it?**
44.2 MB in total: a 17.1 MB acoustic model plus the 27.1 MB Vocos vocoder. Junco and Goldfinch share the vocoder, so
both voices together take 61.3 MB.

**Can it run on a Raspberry Pi or other ARM boards?**
It runs on 64-bit ARM Linux wherever `onnxruntime` and `espeakng-loader` publish wheels, including 64-bit Raspberry
Pi OS.

**Can I stream audio for low latency?**
Yes. `tts.stream(text)` yields audio sentence by sentence, so playback can start before the whole text is done.

**Can I use it without Python?**
The ONNX files run in any ONNX Runtime binding. The text front end (espeak-ng phonemes) and the small loop around the
three models live in the [`flockspeech`](https://github.com/flockspeech/flockspeech) package, which you can port.

**Does it support voice cloning or other languages?**
No. Junco is one English (US) voice. For a casual, conversational voice, use
[Goldfinch](https://huggingface.co/flockspeech/goldfinch).

**Can I use it commercially?**
Yes. The weights are Apache-2.0 and the vocoder is MIT. The phonemiser (espeak-ng, phonemizer) is GPL-3.0 and is
installed separately; see the License section below.

## License

- **Junco weights and configuration:** [Apache License 2.0](LICENSE). Commercial use, modification and redistribution
  are allowed.
- **Vocos vocoder weights** (`onnx/vocoder.onnx`): MIT, © 2023 Charactr Inc. See `onnx/LICENSE-vocos`.
- **Training data:** LJ Speech, public domain.

**Run-time dependency.** Text-to-phoneme conversion uses espeak-ng and phonemizer. Both are GPL-3.0 and are **not**
included in this repository. The `flockspeech` package installs them as separate dependencies. If you plan to ship
Junco inside a closed-source product, review their licence terms, or replace the phonemiser with one under a licence
that suits you. See [NOTICE](NOTICE).

## Citation

```bibtex
@misc{flockspeech2026junco,
  title        = {Junco: a small flow-matching text-to-speech model for on-device use},
  author       = {{Flock Speech}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/flockspeech/junco}},
  note         = {Version 0.1}
}
```

## Links

- Python package: [pypi.org/project/flockspeech](https://pypi.org/project/flockspeech/)
- Source code: [github.com/flockspeech/flockspeech](https://github.com/flockspeech/flockspeech)
- Conversational voice: [flockspeech/goldfinch](https://huggingface.co/flockspeech/goldfinch)

## Acknowledgements

Junco builds on ideas from Matcha-TTS (conditional flow matching for TTS), Glow-TTS (Monotonic Alignment Search),
F5-TTS (sway sampling), ZipVoice and SupertonicTTS (efficient decoder design), and Vocos (vocoder). We thank the authors
of these works and of LJ Speech.