Kestrel is a distilled, frame-rate text-to-speech engine for Apple Silicon (MLX), created by
compressing and re-architecting Kokoro-82M for
single-voice audiobook rendering. Named after the falcon: small, very fast, and precise in timing.
~10 M active parameters (vs Kokoro's 82 M) — the vocoder never leaves frame rate
(hop 300 @ 24 kHz), and a whole chapter runs as a handful of batched, compiled GPU stages.
This is the final state of the project. The research program that produced it ran 113
recorded experiment cycles (every one logged with a pre-registered prediction, a falsifier, and
a KEEP/KILL verdict) and is now closed. The complete decision trail is in the GitHub repo.
Speed (M2 MacBook, 16 GB; 163 s reference chapter; warm, median of 5)
preset
wall clock
real-time factor
vs stock Kokoro
student-fast
0.271 s
×622
~53× faster
student-fast-mask (legacy head)
0.251 s
×671
~57× faster
student
1.117 s
×146
~12× faster
stock Kokoro-82M (MLX)
13.7 s
×11.9
1×
Quality — measured, stated plainly
Reference-free perceptual instruments (higher is better; the teacher is the ceiling being
distilled toward):
worst case only on adversarial punctuation text; 0–8 % on narration
MCD (DTW)
13.89 ❌
control bar 3.98 — the spectral-similarity gate fails
What this means in practice: same narrator, same words, fully intelligible (ASR at parity
with the teacher), and the final vocoder scores within ~0.2–0.3 MOS of the teacher on three
independent perceptual predictors — but frame-wise spectral similarity to the teacher (MCD) is
far from the control bar, and we do not claim parity in a direct A/B. Use Kestrel where
throughput matters and a slight texture difference is acceptable.
Architecture (final)
text ──FastG2P (~4 ms/chapter, misaki-equivalent)──▶ phonemes
├─ student: batch-exact Kokoro phoneme path (masked fp16 BiLSTM scans)
│ → bit-exact durations + text/duration-encoder features
└─ student-fast: fully distilled phoneme encoder + duration regressor
──▶ F0/N student (ConvNeXt-1d on duration-encoder features)
──▶ decode student (frame-rate ConvNeXt-1d, L1-distilled from the teacher's decode blocks)
──▶ SFNoiseHead vocoder (source-filter): a true-sinusoid, alias-gated harmonic excitation
built in the time domain from exact F0 phase, shaped by a bounded log-magnitude+phase
filter, plus an additive noise envelope → single iSTFT, hop 300
The source-filter head replaced the original MaskHead (spectral-mask-over-template) after the
research loop measured MaskHead's representational ceiling; it was then trained adversarially
against a 7-lens discriminator ensemble (HiFi-GAN MPD+MSD + multi-resolution log-spectrogram
lenses) with gradient-informed lens weighting. The legacy MaskHead remains available as the
student-fast-mask preset.
Files
file
what
kestrel_sf_lw58k.safetensors
SFNoiseHead vocoder — the final default head (cycle 113)
kestrel_maskhead.safetensors
legacy MaskHead vocoder (student-fast-mask)
kestrel_decode.safetensors
decode-blocks student
kestrel_f0n.safetensors
F0 / energy student (used by student)
kestrel_prosody.safetensors
phoneme encoder + durations (used by student-fast)
student-fast is self-contained. student additionally loads the original Kokoro-82M weights
(fetched from the Hub on first run) for its bit-exact phoneme/duration path.
Training
Distilled entirely on one M2 MacBook (16 GB) from the frozen fp32 Kokoro-82M teacher:
~5 h of captured teacher renders and intermediate features, multi-resolution STFT + mel +
cepstral losses, then adversarial training against a 7-lens discriminator ensemble (MPD + MSD +
multi-resolution log-spectrogram lenses, gradient-balanced) with best-of-snapshots selection
under a three-instrument battery (UTMOS + NISQA + DNSMOS, two-instrument agreement required for
any shipped claim). Single voice (af_heart), English only.
Limitations
One voice, English only; other languages/voices not trained or evaluated.
The frame-wise spectral gate fails (MCD 13.89 vs the 3.98 control bar) — texture is not
claimed identical to the teacher.
student-fast duration drift is 4.97 % mean but up to 50.3 % on adversarial
punctuation-dense text (measured cause: the distilled duration path lacks the teacher's
style-conditioned response; seven experiment cycles closed this as inherent to the 80 fps
representation).
speed != 1.0 is supported (matching the teacher's duration/speed semantics); slowing
below 1.0 is the weaker direction (+0.38 dB MCD at 0.8×).
The included G2P lookup tables are tuned to an audiobook corpus; out-of-domain words fall
back to espeak (slower first pass, then memoized).
Provenance & license
Apache-2.0, matching the Kokoro-82M teacher from which these weights are distilled.
Full engineering report, the complete 113-cycle decision trail with every measured dead end,
and an independent "did we cheat?" audit are in the GitHub repo's docs/ and
experiments/LEDGER.md.
