Hi, I'm Dusk
I'm a neural voice, and I'm still growing up. I speak English (US), I'm free and open source, and I live inside Piper, a text-to-speech system that runs completely offline, right on your own device.
The person who made me listens to a screen reader all day, every day, so I'm being made to be a voice that actually feels good to listen to. Here's who I'm trying to be:
- Clear. Clean, easy to understand, no mumbling.
- Lively. Real ups and downs in my pitch, not a flat drone.
- Quick. Still crisp when you crank my speech rate way up.
Where I am right now: still in training! I'm only 25 of my 110 rounds of training (epochs) in, and I still sound a bit raspy. I get smoother every session, and my maker decided to wait until I'm really good before I come home with anyone.
Download me
I'm not ready to come home with you yet. My maker decided to wait until I'm really good, with no rush and no deadline. When I'm finished, my download will be right here.
Until then, I'll be right here practicing. When I'm released, the Install section walks you through every step.
What I am
I'm one voice model for Piper. Like every Piper voice, I come as two files that always go together:
- My model file (it ends in
.onnx). This is my trained neural network, the actual voice. It's the big one. - My settings file (it ends in
.onnx.json). A small text file with my name, my language, my sample rate, my default speed, and the list of speech sounds I know.
Apps need both of my files, with matching names, in the same place. When I'm all grown up, my files will be called en_US-dusk-high.onnx and en_US-dusk-high.onnx.json, and apps will just show me as "Dusk".
How I talk
I'm built on a kind of neural network called VITS, like all Piper voices. Getting from your text to my voice takes three steps:
- Your text becomes speech sounds. A free tool called espeak-ng turns the text into phonemes, the tiny individual sounds of speech. For example, my name, "dusk", becomes four sounds: d, uh, s, k. That's how I know how to pronounce words.
- Speech sounds become audio. Those phonemes go into my neural network (the
.onnxfile). I work out how long each sound should last and what it should sound like, and then I create the audio directly. - Out comes my voice. You get audio at 22,050 samples per second (22,050 Hz), which your app plays or your screen reader speaks.
All of this happens on your own device, so once I'm installed, I don't need the internet at all.
My speech settings
Three settings inside my .onnx.json file shape how I talk:
- length_scale is my speed. A bigger number makes me slower and a smaller number makes me faster. My default pace is about 144 words per minute, and your screen reader's rate slider can still take me faster or slower from there.
- noise_scale is how much variation goes into my voice itself. Higher makes me more expressive, but I can get shaky. Lower makes me steadier and smoother.
- noise_w is how much variation goes into my timing and rhythm. Higher makes me sound less robotic, lower makes me more even.
How I'm learning to speak
The recordings I learn from
I learned to speak from about 3.5 hours of Celine Major's recordings. She's reader 11697 in the Hi-Fi TTS dataset, which is licensed CC BY 4.0, and her recordings come from LibriVox audiobooks.
I don't learn from every bit of her audio. My training notebook measures all of her clips and picks about 3,069 of them (around 3.5 hours) that are:
- Lively. The clips with the widest pitch range, where her voice moves up and down the most.
- Clean. The noisiest clips, the odd character voices, and the badly aligned clips all get left out.
The clips that make the cut are converted to 22,050 Hz and evened out in loudness before I study them.
Where I started from
Learning to talk from zero would take way too long on free computers. So I started from Piper's public-domain LJ Speech "high" voice, which already knew how to speak English, and my training teaches it to sound like me instead. That's called fine-tuning, or warm-starting.
Where I train, and for how long
- I train on Google Colab's free T4 GPU, using the notebook in my repo.
- I learn from 12 clips at a time (a batch size of 12). My maker tried 16, but that ran the T4 out of memory.
- The plan is 110 epochs. One epoch is one pass through all of my clips, and all 110 add up to about 19 hours of GPU time.
- Colab disconnects free sessions, so I train in sessions. My checkpoints (saved snapshots of me) go to Google Drive, and each new session picks up from the newest one.
- Colab's free GPU time has a limit. When it runs out, I wait until Google gives the GPU back, so finishing me takes several days. Good things take time.
My versions: Compact, Standard and Enhanced
Every version is the same me. They just differ in size and in how long I've trained.
Dusk Compact
- Made for the iPhone app Piper - Neural TTS.
- Half precision (fp16), about 58 MB.
- I also come split into two parts, an encoder and a decoder, so I use less memory. iOS gives voice apps very little memory, and a full-size voice can get cut off on long text, which makes VoiceOver fall back to its default voice. Compact keeps that from happening to me.
- Files:
dusk-compact-ios.zip, with the half-sizeen_US-dusk-high.onnx, its.onnx.json, plusen_US-dusk-high.encoder.onnxanden_US-dusk-high.decoder.onnx.
Dusk Standard
- Made for NVDA, Android, Piper on a computer, and Home Assistant.
- Full precision (fp32), about 110 MB.
- Files:
dusk-standard.zip, with the full-sizeen_US-dusk-high.onnxand its.onnx.json.
Dusk Enhanced
- Coming later.
- The same me, trained longer.
How my Compact files are made
My Compact files come from a half-precision conversion plus Piper's own encoder and decoder splitter. My maker tested this recipe on the public LJ Speech "high" voice first, and it matched the Piper iPhone app's own published compact files: same inputs and outputs, and identical audio with randomness turned off. Split voices are new in the iPhone app, though, so some app versions might only accept my half-size .onnx and .onnx.json without the two split files.
My previews
Previews are early test versions of me, made from a checkpoint partway through training, so my maker can hear how I'm coming along.
Where my previews went
My first preview came from my epoch 25 checkpoint. I was understandable, but still raspy and a little shaky, and slow to start talking on a phone. So my maker took it down. We'd both rather wait until I'm truly good than hand you a voice that isn't ready. There are no previews to download right now.
Why every preview of me gets its own name
The iPhone app keeps track of voices by the name inside the settings file. If two voices share the same identity, the new one replaces the old one, or the app just keeps the old one loaded. So any preview of me gets its own name and dataset id (for example dusk_preview_1). That way my previews and the finished me can live side by side on one phone.
What's next for me
- Keep training until all 110 epochs are done, then release Dusk Standard and Dusk Compact.
- My next build: blending in more openly licensed voices. The plan is to add recordings from Kristin (public domain, from LibriVox) and HFC Female (licensed CC BY-NC-SA 4.0) alongside Celine Major. The goal is a me with lively pitch, strong clarity and crisp consonants.
- Dusk Enhanced: the same me, trained longer.
- Someday, maybe: I'd love to be on Piper's official voice list. That one's up to the Piper maintainers.
A quick, honest license note about my blended build: if HFC Female's recordings are used, that build's voice files will be CC BY-NC-SA 4.0. That means non-commercial use only, and anything made from them has to use the same license. Builds of me trained only on Celine Major (and public domain data) stay CC BY 4.0.
Install me
My finished files will live on my Dusk releases page on GitHub. Short steps for each app are below, and there are longer notes in my Dusk integration notes.
iPhone (VoiceOver)
When I'm released, you'll use the free, open-source app Piper - Neural TTS on the App Store. It needs iOS 18.
- In Safari, open my releases page and download every file for the version you want: my voice model (the
.onnxfile) and my settings file (the.onnx.jsonfile). - The files land in the Downloads folder in the Files app. That's in On My iPhone or in iCloud Drive, depending on your Safari settings. To make sure I end up on your phone, set Safari's download location to On My iPhone in Settings, under Safari, Downloads.
- Open Piper - Neural TTS, go to Settings, then Import Model from Files.
- Open that Downloads folder, select all of my files in one pick, and tap Install.
- In iPhone Settings, go to Accessibility, VoiceOver, Speech, then Add Voice (or Voice), English, and choose me by name. You can pick me for Spoken Content too.
Tip if my files look dimmed: the import picker only works with real local files in On My iPhone. Files that are still sitting in iCloud Drive, Google Drive, Dropbox or another cloud provider can show up dimmed, or fail to import without saying anything. If that happens, copy them into On My iPhone first (for example: open the file in the cloud app, then Share, Save to Files, On My iPhone). Downloading me straight from Safari skips this problem.
iPhone (VoiceOver), with Dusk Compact
For my phone-sized build, it's almost the same:
- Download
dusk-compact-ios.zipfrom my releases page and unzip it in the Files app. - Make sure the files are real local files in On My iPhone (see the tip above).
- Open Piper - Neural TTS, go to Settings, then Import Model from Files.
- Select
en_US-dusk-high.onnxanden_US-dusk-high.onnx.jsonin one pick. If your app version supports split voices, also selecten_US-dusk-high.encoder.onnxanden_US-dusk-high.decoder.onnx. Then tap Install. - In iPhone Settings, go to Accessibility, VoiceOver, Speech, then Add Voice (or Voice), English, and choose Dusk. You can pick me for Spoken Content too.
Updating to a newer me? After you import it, switch VoiceOver to another voice and back (or restart your phone) so it loads the new one.
NVDA (Windows), with Dusk Standard
- Install the Sonata Neural Voices add-on for NVDA.
- Download
en_US-dusk-high-nvda-sonata.tar.gzfrom my releases page. - In NVDA, open Sonata's voice manager, go to the Installed tab, choose "Install from local file", and pick the file.
- Restart NVDA, then choose Sonata and Dusk in NVDA's speech settings.
Android, with Dusk Standard
- Install SherpaTTS from F-Droid (Android 10 or newer).
- Download
vits-piper-en_US-dusk-high.tar.bz2from my releases page and unpack it. - In SherpaTTS, go to Manage Languages, then Install from SD Card. Use the language code
engand the nameDusk, then picken_US-dusk-high.onnxandtokens.txtfrom the unpacked folder. - In Android Settings, go to Accessibility, Text-to-speech output, and choose SherpaTTS.
Piper on a computer (Windows, Mac, Linux), with Dusk Standard
Put both of my Standard files in one folder, then run these two commands. The last part is me saying hello:
pip install piper-tts
python3 -m piper -m en_US-dusk-high --data-dir /folder/with/both/files -f test.wav -- "Hi, I'm Dusk."
Home Assistant, with Dusk Standard
Put both of my Standard files in /share/piper and restart Home Assistant.
JAWS (Windows): not tested yet
JAWS can use Windows SAPI 5 voices, and the Windows text-to-speech bridge for Piper voices is a SAPI 5 engine that runs Piper voices. Nobody has confirmed I work with JAWS yet, so I'm calling it untested for now. Steps are in my Dusk integration notes.
Train me yourself
Want to raise your own copy of me? My whole training process lives in one notebook: the Dusk Colab notebook.
- Open the notebook in Google Colab.
- In the Runtime menu, choose Change runtime type, pick T4 GPU, then Save.
- In the Runtime menu, choose Run all (Ctrl+F9). Allow Google Drive access when it asks.
- Everything gets saved in one folder in your Google Drive,
Dev/Dusk: my data, checkpoints, samples, logs and the finished voice. The notebook doesn't touch anything else in your Drive. - If Colab disconnects or your free GPU time runs out, no worries. Come back later, connect again and press Run all. It skips whatever's already done and picks up from my newest checkpoint.
- The Preview cell makes an early test version of me from the newest checkpoint and saves it in
Dev/Dusk/export/preview-1. - When my training's done, the finished me shows up in
Dev/Dusk/exportasdusk-standard.zipanddusk-compact-ios.zip.
The Settings cell at the top is the only one you should need to change. And if something breaks, the notebook's last section, "If something goes wrong", covers the common errors, like running out of GPU memory.
Credits
I wouldn't exist without a lot of generous people sharing their work. Thank you, all of you!
- Voice data: "Hi-Fi Multi-Speaker English TTS Dataset" (Hi-Fi TTS) by Evelina Bakhturina, Vitaly Lavrukhin, Boris Ginsburg and Yang Zhang (NVIDIA), 2021. Hi-Fi TTS on OpenSLR. Licensed Creative Commons Attribution 4.0 (CC BY 4.0).
- Reader: Celine Major, Hi-Fi TTS speaker 11697, a LibriVox volunteer reader. Original recordings from LibriVox, public domain.
- Changes to the data: about 3.5 hours of her clips, selected for pitch range and cleanliness, resampled to 22,050 Hz and loudness-normalized.
- Starting model: Piper "ljspeech" high checkpoint by Bryce Beattie (public domain), from the Piper checkpoints collection on Hugging Face, trained on the LJ Speech Dataset by Keith Ito and Linda Johnson (public domain).
- Trained with Piper: Piper's training code on GitHub. Piper's code is GPL-3.0. It's used to train me, and it isn't part of my voice files.
- iPhone app: Piper - Neural TTS by Ihor Shevchuk, source code for the iPhone app.
I'm not endorsed by the reader, NVIDIA, LibriVox, the Piper project or the app's developer. My Dusk model card has the full data notice.
License
- Code (the notebook and scripts): MIT License.
- My voice files (once released): CC BY 4.0, which comes from my training data. Use me, share me and build on me, even in apps. Just keep the credits above with me.
- Blended builds: a build of me that uses HFC Female's recordings will be CC BY-NC-SA 4.0 (non-commercial, share-alike), like I explained in What's next for me.
Thanks so much for stopping by to meet me. I hope I make your day sound a little nicer.