Our voice models

Natural speech
in 44 MB.

Flock makes text-to-speech models small enough for your own computer, each one named after a bird. They run on an ordinary CPU, with no GPU and no internet once they are installed. Junco reads, Goldfinch chats and laughs, and more are on their way.

The Flock Speech team

An American goldfinch perched on the card

Pick a line

Hey, I'm Goldfinch. laugh I'm the chatty one, and, uh, I live on your computer too.

0:05

Spoken by Goldfinch, the conversational voice, on a laptop CPU, ahead of time. Tags like [laugh] become sounds. A version you can type into, running in your browser, is coming next.

per voice, vocoder included
44.2 MB
per voice, vocoder included
of memory while a voice speaks
215 MB
of memory while a voice speaks
Goldfinch chats, Junco reads, one install
2 voices
Goldfinch chats, Junco reads, one install
audio, on an ordinary CPU, offline
24 kHz
audio, on an ordinary CPU, offline

Junco and Goldfinch are out. Install both with pip install flockspeech, or find it on PyPI, with model cards for Junco and Goldfinch on Hugging Face and the code on GitHub. Listen below, then try it on your own computer.

Hear it

Two voices, one tiny package

Goldfinch chats, Junco reads. Every line below was spoken on a laptop CPU, ahead of time, by the same package you install.

Goldfinch

Conversational voice

Relaxed and chatty. It says “uh” and “you know”, and laughs or breathes where you write a tag.

  1. A greeting

    “Hey, I'm Goldfinch. laugh I'm the chatty one, and, uh, I live on your computer too.”

    0:05

  2. A laugh

    “Oh my gosh, laugh I can't believe you actually did that.”

    0:03

  3. Small talk

    “Yeah, you know, I used to work there, but, um, not anymore.”

    0:03

  4. A plan

    “breath Okay, so here's the plan for tomorrow. We meet at nine, and, uh, we'll figure out the rest.”

    0:06
flockspeech "…" --voice goldfinchv0.1 · 11.4M parameters · 44.2 MB

Junco

Reading voice

Clear and neutral, for assistants, screen readers, notifications and narration.

  1. A greeting

    “Hello. I'm Junco, a small voice that lives on your computer, not in the cloud.”

    0:05

  2. A quiet moment

    “The kettle clicked off, and for a moment the whole kitchen went quiet.”

    0:04

  3. Directions

    “Turn left at the old stone bridge, then follow the river until you see the mill.”

    0:05

  4. A question

    “Did you remember to water the plants before you left this morning?”

    0:03
flockspeech "…" --voice juncov0.1 · 11.4M parameters · 44.2 MB

Goldfinch

A voice that laughs where you tell it to

Goldfinch talks the way people do: it says “uh” and “you know”, and it laughs or takes a breath wherever you write a tag. Same size and speed as Junco.

An American goldfinch perched on the card

“That's hilarious. laugh No, really, it is.”

Goldfinch v0.1

0:02

Fillers need no tags: write “uh”, “um” or “you know” as words. Each tag is one symbol the model learned from real conversations, so it decides how the laugh sounds and how long it lasts. Model card

$ flockspeech "That's hilarious. [laugh] No, really." --voice goldfinch --play

Sound tags

  • [laugh] a short laughgood
  • [breath] an audible breathgood
  • [throat] clearing the throatfair
  • [lipsmack] a lip smackfair
  • [sigh] a sighweak

Good to know

  • Put a tag where the sound belongs, between words or at the start of a sentence.
  • Try another seed if a laugh sounds off: each take is a little different.
  • Junco reads the same text and quietly skips the tags.

The flock

Every voice is a bird

Small birds for small voices. Two are out today; each release gets a name, a version and a label you can check.

A dark-eyed junco perched on Junco’s card

Junco

Available

The first voice. One English speaker, trained from scratch for three hours on a single T4 GPU. It runs on one CPU thread.

0:05

v0.1 · 11.4M parameters · 44.2 MB

An American goldfinch perched on Goldfinch’s card

Goldfinch

Available

The conversational voice. It says “uh” and “you know”, and laughs or breathes where you write [laugh] or [breath]. Same size and speed as Junco.

0:03

v0.1 · 11.4M parameters · 44.2 MB · sound tags

A golden-crowned kinglet perched on Kinglet’s card

Kinglet

Planned

Junco’s weights stored as 8-bit integers, about a quarter of the size.

int8 · for phones and small boards

A common swift flying above Swift’s card

Swift

Planned

A version that needs fewer steps to speak, for the fastest replies. Swifts hardly ever land.

distilled · fewer decoding steps

Run it

One command, on the computer you already have

No GPU, no account, nothing sent anywhere. Everything the voices need comes in the package, so it works offline from the first run. Code on GitHub, model cards for Junco and Goldfinch.

A dark-eyed junco perched on the terminal
TerminalmacOS · Linux · Windows
$ pip install flockspeech
Successfully installed flockspeech-0.2.0
$ flockspeech "Hello there." --play
Junco v0.1: 1.2 s of audio -> out.wav, playing
$ flockspeech "Oh my gosh, [laugh] really?" --voice goldfinch --play
Goldfinch v0.1: 1.9 s of audio -> out.wav, playing
# or from Python
$ python -c "from flockspeech import Junco; Junco().save('Hi.', 'hi.wav')"

Offline

nothing is sent anywhere, no account needed

61.3 MB

both voices installed; they share one vocoder

24 kHz

audio out