We gave our AI a voice in one day. Here's the recipe — and the voice.

The numbers first

Measured on our own hardware, because model cards that only brag are ads.

The trick: the corpus is free if you own the teacher

You can't clone a voice into this model — there's no cloning interface. But you can distill one: have a big TTS voice you like read a few thousand sentences, then teach the tiny model to imitate it. The beautiful part: because you wrote the sentences, every transcript is perfect by construction. No transcription step, no labeling budget. We rendered ~4,400 clips — about five and a half hours of 24kHz audio — from a commercial cloud narrator voice.

The three voices that came after didn't need a vendor at all. They were described, not recorded. One sentence — “a deep, authoritative older male narrator, gravelly texture, slow and weighty, movie-trailer gravitas” — went into a voice-design model on our own hardware, and what came back was Silas. Clara and Pip arrived the same way, each from a single sentence. That description is the entire origin of the voice; everything after it is the same distill.

The step everyone skips is the one that saved us

Before training, every clip went through a gate: speech recognition transcribes it back, and if the audio doesn't say what the script says, it dies. Boring, right?

The race: laptop GPU vs. rented iron

Same corpus, same config, same 50,000 steps — two different machines, timed.

The rental started four hours behind and finished two hours ahead, for roughly the cost of a sandwich. The laptop spent the afternoon at 98% GPU with the fans screaming. Verdict: laptops are for pilots; rent the training runs. Your hardware — and your lap — will thank you.

What transferred, and what didn't

The voice's identity — timbre, warmth, the basic read — transferred remarkably well. What a 9M model can't fully hold is the teacher's timing: the dramatic pause, the phrase-level planning. The organ responsible for timing in this architecture is the size of a walnut. Scrappy reads a little flatter than his teacher, and we're telling you that up front.

Before and after, in their own words

Run him yourself

Everything is Apache-2.0 — weights, samples, and the fine-tuning stack the original release doesn't ship. CPU is all you need.

# get it (weights included — ~40MB)
git clone https://github.com/scrappylabsai/scrappy-voice-1
cd scrappy-voice-1
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# say something
python inference.py --model-dir . --device cpu \
  --text "Hello World. I'm Scrappy." --output hello.wav

Or from Python:

from inference import InflectTTS

tts = InflectTTS(model_dir=".", device="cpu")
tts.save("Write numbers as words for best results.", "out.wav", seed=7)

Knobs: --speed 0.5–2.0, --variation 0.0–1.0 (lower = steadier, higher = livelier), --seed for reproducible takes. English only, one voice — his. Want a different voice in there? The trainer/ directory in the repo is the fine-tuning stack we had to build: corpus prep with phoneme validation, the full warm-start training loop, and candidate export. Bring ~1–5 hours of clean single-speaker audio with verified transcripts, and gate it like we told you.

Two days later: he has a family

One voice proved the pipeline. So we ran it three more times — and then did the thing the architecture was quietly built for: trained all four into a single checkpoint. Each voice is one embedding row, which means four voices cost one model, and the space between any two of them is a real place you can stand.

All four clips above came out of the same 41MB file, on a CPU. The pleasant surprise: pooling four voices' data made the shared model better than the single-voice runs, not worse — the four voices landed at roughly the quality of our best solo run instead of paying for the company.

The demo runs on our own hardware. Drag the blend slider and you are interpolating between two speaker embeddings live — near the ends it sounds like a person, and the deep middle gets strange, because only the four anchors were ever trained. Everything between them is the model improvising.

Solo models, one voice each: Clara · Silas · Pip · Scrappy