[lalalune]

What

1. paradee_load_model memory-maps an all-F32 GGUF.
   • That covers the converter's default and the pinned file. It is mapped read-only and the layers read the weights in place.
   • Until now the loader read the whole file into private memory: 32 MB that stays resident for as long as the voice is loaded.
   • PARADEE_USE_MMAP=0 restores the read. Non-F32 files are still read and widened, and Windows falls back to the read.
2. README: the evaluation of --outtype f16 and of Paradee-8M v1.1 that was asked for, with the numbers below.

Mapping: evidence

• test-paradee-parity against a fresh reference dump (pinned torch 2.14.0 / kokoro 0.9.4 / misaki 0.9.4, paradee 9c8b4d75): OK, mapped and read.
  • The native waveforms are byte-identical between the two modes (cmp, all 5 sentences).
• Also OK: test-paradee-api (mapped, read, and an F16 file through the widening path), test-paradee-phase-lock, and eliza-inference-test-paradee-asr (558 checks).
• TB336FU (4 GB, Android, arm64), paradee-tts on 19.8 s of speech, two runs each:

	peak RssAnon	peak RssFile	RTF
mapped (new default)	170 MB	34 MB (clean)	0.29–0.34
PARADEE_USE_MMAP=0 (old behaviour)	202 MB	3 MB	0.28–0.32

F16: keep F32

Pinned v1.0 GGUF, this directory's parity test, and 40 more sentences (20 assistant prompts and 20 Harvard sentences):

• The sentences were phonemized with tools/misaki-g2p, synthesized with paradee-tts, and transcribed with Phonon-2 (tools/parakeet) and whisper.cpp base.en.
• "Mixed" means an F16 decoder with an F32 BERT and predictor. It was an experiment and is not in the converter.

File	Size (deflate)	Parity test	Durations	WER Phonon-2 / whisper	RTF tablet / M4
F32 (pinned)	32.1 MB (29.7)	pass (end to end 58–76 dB)	–	3.90% / 3.90%	~0.30 / 0.024
--outtype f16	16.2 MB (14.9)	fail (end to end 9.8–23 dB)	1 of 40 changed by 25 ms	3.90% / 3.90%	~0.32 / 0.024
mixed	22.1 MB (20.4)	end to end 56–61 dB, but asr/dec exceed the 1e-3 × RMS stage gate	identical	3.90% / 3.90%	~0.30 / 0.024

• The F16 error is in the BERT and the duration/F0 predictor. With those in F32, the end-to-end SNR rises to 56–61 dB. The low SNR of full F16 is mostly accumulated harmonic phase from F0 errors near 1e-3: the log-mel difference is small and WER is unchanged.
• F16 is not faster, because the runtime computes in F32.
• F16 costs more RAM. It cannot be mapped and is widened at load, which adds 16–22 MB of private memory (measured on the tablet: 218–223 MB peak vs 202 MB for F32 read, and 170 MB for F32 mapped).
• So F32, mapped, is the right file today. F16 would only pay off with layers that read F16 weights in place. The ggml conv1d path already keeps a private F16 copy of its weights (cached_f16_weight, up to ~10 MB for the 3-D weights). That is where a follow-up could save memory.
• The published "8 MB" is upstream's int8 ONNX graph (onnx/paradee_int8.onnx, 9.0 MB). This runtime does not use it.

v1.1: needs a native port, not a pin move

• sahilmahendrakar/Paradee-8M-v1.1 at 330617ad keeps the text side (identical text_side.pt) and the decoder's AdaIN trunk.
• It replaces the iSTFTNet generator and the phase lock with a frame-rate source-filter head (training/scripts/vhead.py, paradee 606e16e1): ConvNeXt blocks at 80 fps shaping the STFT of a 64-harmonic sine bank and of random noise, plus a 3 ms noise burst path.
• The current converter and runtime cannot load it. Its output is also not deterministic (it draws noise), so the parity gate would need a seeded noise input.

Upstream's ONNX graphs, one onnxruntime thread on the M4 Max, the same 40 sentences:

Graph	Size	RTF	WER Phonon-2 / whisper
v1.0 fp32	37.0 MB	0.051	4.16% / 3.90%
v1.0 int8	9.0 MB	0.054	3.90% / 3.12%
v1.1 fp32	34.8 MB	0.010	3.90% / 3.12%
v1.1 int8	8.5 MB	0.011	4.16% / 3.12%

v1.1 is about 5x faster at the same WER (one word in 385 is the resolution here). Upstream reports less buzz (2–4 kHz harmonic contrast 3.9 dB vs 1.3 dB); that was not re-measured here. On the 4 GB tablet, where v1.0 runs at RTF ~0.3, a native v1.1 head is the TTS change worth doing next.

Not covered

• No APK and no in-app run.
• The tablet RTF figures were taken while the device was busy with other work (load average ~17), so treat them as relative.

AI-assisted change.

🤖 Generated with Claude Code