scrappylabsai/scrappy-voice-1

Scrappy — a 9M-param CPU voice, distilled in one day

Scrappy is the voice of ScrappyLabs: a full text-to-speech voice in 9.36M parameters / 37MB, running 10–14× faster than real-time on a plain CPU — no GPU, no cloud, 0.16s load. It's a fine-tune of owensong/Inflect-Micro-v2 (Apache-2.0), trained the same day that model crossed our radar.

Listen: samples/scrappy_intro.wav — Scrappy says hello (generated on a desktop CPU by this exact checkpoint).

🎮 Try him live: interactive demo on Spaces — built by the Hugging Face team on free ZeroGPU. Thanks, HF 🤗

Get it: 🤗 HuggingFace (canonical weights) · 🐙 GitHub (clone-and-run mirror, issues & PRs) · Story: scrappylabs.ai/scrappy-voice

Why this exists

Inflect-Micro-v2 ships as an inference-only release — one fixed voice, no trainer, and the docs say voice replacement is "research use." We wanted to know how true that is, so we treated it as a one-day exercise: reconstruct the training stack, distill a voice we like into it, and publish what we learned. This repo is the result plus the missing piece we had to build — a working fine-tuning setup (see trainer/).

The recipe, end to end:

1. Teacher renders the corpus. A commercial cloud TTS narrator voice generated 4,400 short clips (5.5h @ 24kHz) from a text corpus we control — so every transcript is known by construction.
2. An ASR gate cleans it. Every clip is round-tripped through speech recognition and scored against its transcript. This caught real poison: clips where the teacher API quietly injected a spoken watermark instead of the requested text. Signal checks (clipping/silence/duration) run alongside. 98.6% survived.
3. Warm-start fine-tune. The released model.pth initializes the generator; the posterior encoder and discriminator start fresh (the release runtime ships the full training-side model code — only losses, data loading, and the loop needed writing). Decoder frozen for the first 3k steps, LR 1e-4 → 5e-6 over 50k steps, batch 24, fp32.
4. 50,000 steps ≈ 3 hours on one rented RTX PRO 6000 Blackwell (~$9). A 24GB laptop GPU does the same run in ~6 hours if you'd rather boil a laptop (we don't recommend it).

Usage

Identical to upstream — this is a drop-in checkpoint for the packaged runtime:

from inference import InflectTTS

tts = InflectTTS(model_dir=".", device="cpu")
tts.save("Hi, I'm Scrappy. I run on your CPU.", "out.wav", seed=7)

python inference.py --model-dir . --device cpu --text "Hello from Scrappy." --output out.wav

Notes carried over from upstream: English only, single voice, deterministic seeds, punctuation-aware long-form chunking, speed 0.5–2.0, variation 0.0–1.0. Write numbers out as words for best results.

Fine-tune your own voice (trainer/)

The trainer/ directory contains the training stack upstream deliberately omits:

• prep_filelists.py — phonemizes transcripts with the model's own frontend and validates every symbol against the release inventory (espeak emits out-of-vocabulary diacritics on exotic names — unvalidated, they crash training).
• train_ft.py — the full loop: VITS losses, torchaudio mel transforms (slaney/slaney — no librosa dependency), warm-start loading, decoder freeze schedule, drop-in candidate export every N steps.
• eval_candidate.py — renders fixed prompts from any candidate for A/B listening.

You also need the cython monotonic-alignment kernel from the canonical VITS repo (the release stubs it out) — build it and drop the package into runtime/. Data contract: mono 24kHz clips, one speaker, verified transcripts, 1–5+ hours. Gate your corpus with ASR round-trips — it's the only check that catches audio that says the wrong thing beautifully.

Honest limitations

• Prosody is where distillation loses the most. Timbre and identity transfer well; the teacher's long-range timing instincts (dramatic pauses, phrase-level planning) get averaged. The duration predictor is the smallest organ in a VITS — expect a flatter read than the source voice.
• Slight texture softness vs. a large vocoder remains at close listening.
• Everything upstream says about biases and English-only applies.

Provenance & takedown

The training audio was synthesized by a commercial cloud TTS narrator voice (a synthetic persona — no real person's voice was cloned). If you're a rights holder with a concern, open a discussion on this repo and we'll respond promptly.

Credits

• owensong/Inflect-Micro-v2 — base model, runtime, and an unusually honest set of docs (Apache-2.0)
• VITS (MIT) — architecture lineage + alignment kernel
• Built in a day by ScrappyLabs — we do this kind of thing daily to stay sharp. Bring your own AI; we keep it wrangled.