Silas — the low, slow one. 9M params, still on your CPU.
Silas is the gravitas voice in the
ScrappyLabs voice family:
deep, older, a little gravelly, unhurried — the read you reach for when a line needs weight
behind it. Like the rest of the family he is
9.36M parameters / 37MB, 24kHz, running
10–14× faster than real-time on a plain desktop CPU — no GPU, no cloud. He's a fine-tune
of
owensong/Inflect-Micro-v2 (Apache-2.0).
Listen:
samples/silas_intro.wav — Silas introduces himself
(generated on a desktop CPU by this exact checkpoint).
The family: scrappy ·
silas (you are here) ·
clara ·
pip ·
all four in one 41MB model
(multi-speaker + voice morphing — Silas is speaker id 2 there)
🎮 Try all four voices — and blend between them: interactive demo on Spaces — the Space was built and gifted to us by the Hugging Face team. Thanks, HF 🤗
What this is
A single-voice checkpoint, drop-in for the packaged Inflect-Micro-v2 runtime. Same recipe we
used for Scrappy, pointed at a different voice:
- A teacher renders the corpus. Silas did not exist before this — he was described,
not recorded. A voice-design model running on our own hardware turned one sentence,
"A deep, authoritative older male narrator, gravelly texture, slow and weighty, movie-trailer gravitas.",
into a speaking voice, and that voice then read ~4,400 short clips (~5.5h @ 24kHz) from a
text corpus we control — so every transcript is known by construction.
- An ASR gate cleans it. Every clip is round-tripped through speech recognition and
scored against its transcript (≥0.85 word overlap to survive), with signal checks
(clipping/silence/duration) alongside. 99.3% of Silas's clips passed.
- Warm-start fine-tune. The released
model.pth initializes the generator; posterior
encoder and discriminator start fresh. Decoder frozen for the first 3k steps,
LR 1e-4 → 5e-6 over 50,000 steps, batch 24, fp32, on one RTX PRO 6000 Blackwell.
Final mel loss 18.1 — the lowest of the three new solo voices, which tracks: a slow,
steady, low-variance read is the easiest thing in the world for a 9M-param model to fit.
- Measured, not vibed. On held-out intro text, ASR word-overlap against the intended
transcript came back 1.00 for this solo checkpoint.
Usage
Identical to upstream — nothing special to know:
1from inference import InflectTTS
2
3tts = InflectTTS(model_dir=".", device="cpu")
4tts.save("Some things are worth saying slowly.", "out.wav", seed=7)
1python inference.py --model-dir . --device cpu \
2 --text "Some things are worth saying slowly." --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. Silas already reads slow; if you push speed below ~0.9 he starts
to drag rather than deepen.
If you want Silas alongside the other three voices — or blended with them — take the
family model instead. It's the
41MB, holds all four, and adds a
--blend flag.
The
trainer/ directory ships the fine-tuning stack upstream omits (filelist prep with symbol
validation, the full VITS training loop, candidate export and eval) — same code that produced
this checkpoint, if you want to distill a voice of your own. See the
Scrappy card for the walkthrough.
Known quirks
- espeak-ng cannot pronounce his name. The frontend renders "Silas" as /siːləz/ —
SEE-laz. Our synthesis inputs respell it "Sighlus", which comes out /saɪləs/ correctly.
If you're writing his name into a prompt and it sounds wrong, that's why: respell it
phonetically. This is a frontend quirk, not a checkpoint defect, and it's the sort of thing
that bites every name a grapheme-to-phoneme rulebook has never seen.
- 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. For a voice whose whole job is
timing, expect to earn some of the drama back with punctuation and explicit line breaks.
- The gravel is texture, and texture is what a small vocoder softens. At close listening
the rasp is smoother than the source.
- Everything upstream says about English-only operation and model biases applies.
Provenance & takedown
The training audio was synthesized by a voice-design model from a written description — a
synthetic persona that never existed until we described it. No real person's voice was cloned, and please don't use this stack to clone one
without their consent. 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
- The Hugging Face team, who built the family's first demo Space on ZeroGPU and handed us
the keys 🤗
- Built by ScrappyLabs. Bring your own AI; we keep it wrangled.
Trained with: scrappylabsai/inflect-trainer — the fine-tuning stack (single-voice, multi-speaker, and the corpus QC gate), open source.
