PiVo-1.4M / REPRODUCE.md
mlboydaisuke's picture
PiVo-1.4M on LiteRT: day-0 files
042f389 verified
|
Raw History Blame Contribute Delete
23.8 kB

Reproduce and check

This page covers how the LiteRT files in this repository were made, what they were compared against, the bars they had to clear, and the results. The checks ran on 2026-11-08. A script printed every number on this page from the result files of those checks; no number was typed by hand.

The reference

The reference is the author's code (host/pivo/: empero-ai/PiVo-1.4M at revision 515d84fc3420da6ae138f1b81017fcbc24d622a8) running on ONNX Runtime 1.31.1. It used the CPU with one intra-op thread and the default graph optimization, with Python 3.12.13 and numpy 2.5.4, on an Apple M4 Max under macOS 27.0. Two runs in separate processes gave bit-identical audio on all 54 utterances. examples/say.py --check builds the same reference for one sentence.

The test set is the 27 sentences of conversion/check_texts.json in both voices, 54 utterances in all. The set was fixed before any model ran.

The floor is the reference's own spread: the same ONNX graphs run under other ONNX Runtime settings. There were three such runs: a second run with the same settings, a run with 4 intra-op threads, and a run with all graph optimizations off (1 thread).

The bars

The bars were written on 2026-11-08 at 17:19 JST. At that point the reference and floor numbers were in, and no LiteRT output had been compared with the reference yet. The bars were not changed afterwards.

  • Line A: per graph, for the float32 forms (CPU fp32, and GPU at FP32 precision). Each graph is swapped in alone and fed the reference's recorded inputs. On every call and every output, the largest absolute difference must be at most 1e-4 × the largest absolute value of that output in the reference over the 54 utterances. For float16 weights, float16 compute (GPU default precision, NPU) and int8, the per-graph difference is recorded but not judged.
  • Line B: end to end, for every form. Each form synthesizes all 54 utterances, with the Python host on the Mac and the C++ runner on the phones.
    • B1, words: Whisper large-v3-turbo, with the weights' SHA-256 fixed, greedy decoding, temperature 0 and no fallback. The form's corpus WER must be at most the reference's + 1.0 point, and at most 3 of the 54 transcripts may differ from the reference's transcripts.
    • B2, speaker: the ECAPA cosine (version fixed) between the form's audio and the reference audio must be ≥ 0.95 on every utterance.
    • B3, HNR over voiced frames: the absolute difference must be ≤ 1.0 dB on every utterance (floor: 0.013 dB).
    • B4, share of the 2-8 kHz band: the absolute difference must be ≤ 1.5 dB on every utterance (floor: 0.0027 dB).
    • B5, log-mel correlation: must be ≥ 0.99 on every utterance. An utterance with a different length is aligned with DTW first (floor: 0.99982, from one pitch-period flip).
    • B6, length (a bar for the float32 forms, recorded for the others): every utterance must have the reference's sample count. The only allowed exceptions are token-length flips where the reference's unrounded value lies within 1e-3 of the rounding boundary (in log-duration units), and each one is listed.
  • Line C: integer decisions, for the float32 forms. Every pitch period and token length that differs from the reference must lie within 1e-4 of the rounding boundary in the reference (in graph-output units). A flip far from the boundary counts as a defect. For the float16 and int8 forms, flips are counted but not judged.
  • Recorded but not judged: the end-to-end waveform correlation and largest difference, the flip counts, and the time to first audio. The waveform numbers are noisy because of the sample-level recursion: one pitch-period flip of 1 ulp is enough to drop the waveform correlation to 0.9885.
  • The phones' audio is pulled to the Mac and scored there against the same bars. On the Mac, the runner's audio is bit-identical to the Python host's (see "Checks of this repository").

How the files were made

conversion/README.md has the details. Google's LiteRT converters take PyTorch, TensorFlow and JAX models, and none of them takes ONNX. So conversion/pivo_lt.py rewrites the four ONNX graphs in PyTorch, op for op, reading every weight from the ONNX initializers, and litert-torch 0.10.0 converts the result. The rewritten graphs differ from the ONNX graphs in a few ways: static shapes, float32 inputs in place of integer gathers, RoPE tables computed by the host, and a rank-3 feature head in the decoder. conversion/check_graphs.py checks line A.

Checks of this repository

On 2026-11-08, on a copy of this repository in an empty directory (Python 3.12.13, host/requirements.txt and conversion/requirements.txt installed into new venvs):

  • host/pivo_litert.py on the 54 utterances, CPU, 1 thread, with both vocoder_step and vocoder_frame: bit-identical to the audio of the Python host loop scored in the results below. host/pivo_check.py gives the same B3-B5 numbers as the scorer, on three forms.
  • conversion/check_graphs.py --texts conversion/check_texts.json (line A, CPU): every one of the 20 graph outputs is within 1e-4 of its largest value (yes); the worst is vocoder_frame.gru3_out, at 9.92e-06. The numbers are the same as those of the line A check behind the results below.
  • conversion/build_graphs.py --fp32-only from the author's ONNX files: 9 of 9 files byte-identical to this repository's.
  • The runner on the Mac against the Python host, the same files, CPU 1 thread (frame: 54 of 54 bit-identical; step: 54 of 54 bit-identical).
  • runner/build.sh with the kit from runner/fetch_kit.sh: the macOS and Android arm64 binaries are byte-identical to the binaries behind the results; the Mac binary's host arithmetic (phase features, RoPE tables, pitch periods, durations: 272,978 values) against NumPy 2.5.4: 0 values differ in any bit.

Scorers

  • B1: openai-whisper large-v3-turbo, weights SHA-256 aff26ae408abcba5fbf8813c21e62b0941638c5f6eebfb145be0c9839262a19a, options {"language": "en", "task": "transcribe", "fp16": false, "temperature": 0.0, "condition_on_previous_text": false, "verbose": null}.
  • B2: speechbrain/spkrec-ecapa-voxceleb at revision 0f99f2d0ebe89ac095bcc5903c4dd8f72b367286, SHA-256 0575cb64845e6b9a10db9bcb74d5ac32b326b8dc90352671d345e2ee3d0126a2.
  • Packages: python 3.12.13, numpy 2.5.4, torch 2.15.0, soundfile 0.14.0, openai-whisper 20250625, speechbrain 1.1.2, jiwer 4.0.0, torchaudio 2.11.1, scipy 1.18.1.
  • The reference: 8 word errors in 1100 words (WER 0.0073), in 4 of 54 utterances.

Results on the Mac

The Python host (host/pivo_litert.py) on the 54 utterances, on an Apple M4 Max with ai-edge-litert 2.3.0. The reference's WER is 0.0073. (n) = the number of utterances that miss that bar. Line A applies to the float32 forms only, and B6 and C are bars only for those forms; elsewhere they are counted, not judged.

Form Files Line A, worst ratio B1 WER (transcripts differing) B2 cosine min (fail) B3 HNR max diff, dB (fail) B4 2-8 kHz max diff, dB (fail) B5 log-mel corr min (fail) B6 lengths differing C flips: length / period Verdict
CPU (XNNPACK), vocoder_step four times per frame this repository's 9.92e-06 0.0073 (0) 0.999835 (0) 0.0130 (0) 0.0027 (0) 0.999822 (0) 0 0 / 1 PASS
CPU (XNNPACK), vocoder_frame this repository's 9.92e-06 0.0073 (0) 0.999835 (0) 0.0130 (0) 0.0027 (0) 0.999822 (0) 0 0 / 1 PASS
GPU (Metal) at FP32 precision this repository's 1.86e-05 0.0073 (0) 0.999996 (0) 0.0004 (0) 0.0012 (0) 0.999999 (0) 0 0 / 0 PASS
CPU (XNNPACK), the NPU forms _npu_fp32 decoder, encoder, vocoder_frame, vocoder_step; the rest this repository's 9.53e-06 0.0073 (0) 0.999835 (0) 0.0131 (0) 0.0027 (0) 0.999822 (0) 0 0 / 1 PASS
GPU (Metal) at its default precision (FP16) this repository's recorded 0.0073 (0) 0.921743 (4) 0.2344 (0) 0.3020 (0) 0.984032 (6) 29 62 / 590 FAIL
encoder and decoder on GPU (Metal) at FP16, conditioning and vocoder on CPU this repository's recorded 0.0073 (0) 0.928751 (3) 0.2316 (0) 0.3181 (0) 0.984083 (6) 29 62 / 590 FAIL
CPU (XNNPACK), float16 weights _wfp16 conditioning, decoder, encoder, vocoder_frame, vocoder_step recorded 0.0073 (0) 0.964115 (0) 0.1384 (0) 0.1552 (0) 0.988350 (1) 5 6 / 188 FAIL
CPU (XNNPACK), int8 weights _int8 conditioning, decoder, encoder, vocoder_frame, vocoder_step recorded 0.0073 (0) 0.897225 (15) 0.5306 (0) 0.5117 (0) 0.922143 (45) 41 289 / 1489 FAIL

With an earlier decoder (rank-2 feature head), the GPU (Metal) at FP32 precision failed every line: B1 WER 1.2936, with 54 of 54 transcripts different, and B5 min 0.2599. The CPU passed on the same file (PASS). The decoder in this repository writes the head with a rank-3 output, and its GPU row above passes.

What changes in the forms that fail

Form Transcripts differing WER B2 cosine min (fail) B5 log-mel corr min (fail) Lengths differing Token-length flips Pitch-period flips
GPU (Metal) at its default precision (FP16) 0 of 54 0.0073 (reference 0.0073) 0.921743 (4) 0.984032 (6) 29 of 54 62 590
encoder and decoder on GPU (Metal) at FP16, conditioning and vocoder on CPU 0 of 54 0.0073 (reference 0.0073) 0.928751 (3) 0.984083 (6) 29 of 54 62 590
CPU (XNNPACK), float16 weights 0 of 54 0.0073 (reference 0.0073) 0.964115 (0) 0.988350 (1) 5 of 54 6 188
CPU (XNNPACK), int8 weights 0 of 54 0.0073 (reference 0.0073) 0.897225 (15) 0.922143 (45) 41 of 54 289 1489

In every one of these forms the words stay the same (B1); what moves is the voice and the length. With FP16 on the encoder and decoder only, the flips and the failing utterances are the same as with FP16 on every graph (62 token-length and 590 pitch-period flips; 6 utterances under the B5 bar). So float16 in the encoder and decoder moves integer token lengths and pitch periods across the rounding boundary, and the audio follows.

Speed on the Mac

The C++ runner (runner/, LiteRT 2.3.0 C API) on an Apple M4 Max, macOS 27.0, over the 54 utterances. RTF is compute time divided by audio time: pooled over all utterances, and the median of the per-utterance values. The last column gives the seconds of the row during which other jobs were running on the same Mac.

Backend CPU threads Vocoder Files RTF pooled RTF median First audio, median (ms) Other work during the row (s)
CPU (XNNPACK) 1 frame an earlier decoder (rank-2 feature head); the rest this repository's 0.0170 0.0169 0.86 0
CPU (XNNPACK) 4 frame an earlier decoder (rank-2 feature head); the rest this repository's 0.0388 0.0211 1.55 0
CPU (XNNPACK) 1 step an earlier decoder (rank-2 feature head); the rest this repository's 0.0143 0.0135 0.83 0
CPU (XNNPACK) 1 frame _wfp16 conditioning, decoder, encoder, vocoder_frame 0.0171 0.0170 0.86 0
CPU (XNNPACK) 1 frame _int8 conditioning, decoder, encoder, vocoder_frame 0.0093 0.0090 1.04 0
CPU (XNNPACK) 1 frame _npu_fp32 decoder, encoder, vocoder_frame; the rest this repository's 0.0190 0.0187 0.93 0
GPU (Metal), default precision (FP16) - frame an earlier decoder (rank-2 feature head); the rest this repository's 0.0840 0.0698 2.45 0
GPU (Metal), FP32 precision - frame an earlier decoder (rank-2 feature head); the rest this repository's 0.0729 0.0696 2.35 0
GPU (Metal), default precision (FP16) - step an earlier decoder (rank-2 feature head); the rest this repository's 0.1561 0.1496 3.99 34
GPU (Metal), default precision (FP16) - frame this repository's 0.0729 0.0694 2.45 0
GPU (Metal), FP32 precision - frame this repository's 0.0871 0.0700 2.33 0

The reference on the same Mac (the author's code, ONNX Runtime 1.31.1, Python 3.12.13):

Intra-op threads Pass RTF pooled Other work during the row (s)
1 1 0.0274 0
1 2 0.0268 0
4 1 0.0309 0
4 2 0.0282 0

Results on the phones

The C++ runner (runner/pivo_runner.cc) on each phone over the 54 utterances. The audio is pulled to the Mac and scored there against the same reference and bars. First audio = from the start of an utterance to its first samples, warm. (n) = the number of utterances that miss that bar.

Phone Run (JST) Where the graphs ran CPU threads Files RTF pooled RTF median First audio, median (ms) Verdict Details
Galaxy S26 20:11 NPU (Qualcomm HTP), every graph (first run: compiles and writes the NPU cache) 1 _npu_fp32 decoder, encoder, vocoder_frame; the rest this repository's 0.0749 0.0751 4.434 FAIL (B3-B5) B5 min 0.955982 (2); 1 lengths differ; flips 8 / 69
Galaxy S26 20:12 NPU (Qualcomm HTP), every graph (with the NPU cache of the run before) 1 _npu_fp32 decoder, encoder, vocoder_frame; the rest this repository's 0.0763 0.0750 2.628 FAIL B1 0 transcripts differ; B2 min 0.850222 (20); B5 min 0.942459 (27); 36 lengths differ; flips 96 / 780
Galaxy S26 20:13 encoder, decoder, conditioning: CPU (XNNPACK); vocoder: NPU (Qualcomm HTP) (first run: compiles and writes the NPU cache) 4 an earlier decoder (rank-2 feature head); _npu_fp32 vocoder_frame; the rest this repository's 0.0675 0.0675 4.556 B3-B5 pass; B1, B2 to come B5 min 0.993591 (0); 0 lengths differ; flips 0 / 0
Galaxy S26 20:14 encoder, decoder, conditioning: CPU (XNNPACK); vocoder: NPU (Qualcomm HTP) (with the NPU cache of the run before) 4 an earlier decoder (rank-2 feature head); _npu_fp32 vocoder_frame; the rest this repository's 0.0687 0.0674 2.984 FAIL B1 0 transcripts differ; B2 min 0.936823 (5); B5 min 0.988619 (1); 0 lengths differ; flips 0 / 0
Galaxy S26 21:20 CPU (XNNPACK), every graph 1 this repository's 0.0275 0.0252 1.110 PASS B1 0 transcripts differ; B2 min 0.999991 (0); B5 min 0.999997 (0); 0 lengths differ; flips 0 / 0
Galaxy S26 21:22 CPU (XNNPACK), every graph 4 this repository's 0.0357 0.0314 1.531 PASS B1 0 transcripts differ; B2 min 0.999991 (0); B5 min 0.999997 (0); 0 lengths differ; flips 0 / 0
Galaxy S26 21:23 CPU (XNNPACK), every graph 1 this repository's 0.0298 0.0265 1.115 PASS B1 0 transcripts differ; B2 min 0.999991 (0); B5 min 0.999997 (0); 0 lengths differ; flips 0 / 0
Galaxy S26 21:24 GPU (OpenCL) FP16, every graph 1 this repository's 0.1952 0.2060 5.592 FAIL (B3-B5) B5 min 0.930992 (16); 29 lengths differ; flips 76 / 628
Galaxy S26 21:26 GPU (OpenCL) FP32, every graph 1 this repository's 0.2111 0.2196 6.218 B3-B5 pass; B1, B2 to come B5 min 0.999822 (0); 0 lengths differ; flips 0 / 1
Galaxy S26 21:28 encoder, decoder: GPU (OpenCL) FP16; vocoder, conditioning: CPU (XNNPACK) 4 this repository's 0.0410 0.0407 3.105 FAIL (B3-B5) B5 min 0.930998 (15); 29 lengths differ; flips 76 / 628
Galaxy S26 20:25 CPU (XNNPACK), every graph 4 an earlier decoder (rank-2 feature head); the rest this repository's 0.0307 0.0287 1.468 PASS B1 0 transcripts differ; B2 min 0.999998 (0); B5 min 0.999999 (0); 0 lengths differ; flips 0 / 0
Galaxy S26 20:27 CPU (XNNPACK), every graph 1 an earlier decoder (rank-2 feature head); the rest this repository's 0.0254 0.0240 0.974 PASS B1 0 transcripts differ; B2 min 0.999998 (0); B5 min 0.999999 (0); 0 lengths differ; flips 0 / 0
Galaxy S26 21:29 CPU (XNNPACK), every graph 1 this repository's 0.0301 0.0271 1.152 PASS B1 0 transcripts differ; B2 min 0.999991 (0); B5 min 0.999997 (0); 0 lengths differ; flips 0 / 0
Galaxy S26 21:31 CPU (XNNPACK), every graph 1 this repository's 0.0286 0.0260 1.128 PASS B1 0 transcripts differ; B2 min 0.999991 (0); B5 min 0.999997 (0); 0 lengths differ; flips 0 / 0
Galaxy S26 21:31 GPU (OpenCL) FP16, every graph 1 this repository's 0.5330 0.5300 8.320 FAIL (B3-B5) B5 min 0.930992 (16); 29 lengths differ; flips 76 / 628
Galaxy S26 21:35 CPU (XNNPACK), every graph 4 this repository's 0.0390 0.0324 1.586 B3-B5 pass; B1, B2 to come B5 min 0.999997 (0); 0 lengths differ; flips 0 / 0
Galaxy S26 20:57 CPU (XNNPACK), every graph 4 _wfp16 conditioning, decoder, encoder, vocoder_frame 0.0356 0.0306 1.624 FAIL B1 0 transcripts differ; B2 min 0.965478 (0); B5 min 0.988314 (1); 5 lengths differ; flips 6 / 188
Galaxy S26 20:59 CPU (XNNPACK), every graph 4 _int8 conditioning, decoder, encoder, vocoder_frame 0.0149 0.0145 1.190 FAIL B1 0 transcripts differ; B2 min 0.904737 (15); B5 min 0.922159 (46); 41 lengths differ; flips 289 / 1493
Galaxy S26 21:36 NPU (Qualcomm HTP), every graph (first run: compiles and writes the NPU cache) 1 _npu_fp32 decoder, encoder, vocoder_frame; the rest this repository's 0.0747 0.0750 4.302 FAIL (B3-B5) B5 min 0.955983 (2); 1 lengths differ; flips 8 / 73
Galaxy S26 21:37 NPU (Qualcomm HTP), every graph (with the NPU cache of the run before) 1 _npu_fp32 decoder, encoder, vocoder_frame; the rest this repository's 0.0751 0.0748 2.639 FAIL (B3-B5) B5 min 0.942576 (28); 36 lengths differ; flips 96 / 771
Galaxy S26 21:38 encoder, decoder: NPU (Qualcomm HTP); vocoder, conditioning: CPU (XNNPACK) (first run: compiles and writes the NPU cache) 4 _npu_fp32 decoder, encoder; the rest this repository's 0.0380 0.0383 3.854 FAIL (B3-B5) B5 min 0.958339 (2); 1 lengths differ; flips 8 / 73
Galaxy S26 21:39 encoder, decoder: NPU (Qualcomm HTP); vocoder, conditioning: CPU (XNNPACK) (with the NPU cache of the run before) 4 _npu_fp32 decoder, encoder; the rest this repository's 0.0384 0.0349 2.240 FAIL (B3-B5) B5 min 0.942467 (22); 36 lengths differ; flips 96 / 771
Galaxy S26 21:40 encoder: NPU (Qualcomm HTP); decoder, vocoder, conditioning: CPU (XNNPACK) (first run: compiles and writes the NPU cache) 4 _npu_fp32 encoder; the rest this repository's 0.0484 0.0423 2.143 FAIL (B3-B5) B5 min 0.942311 (22); 36 lengths differ; flips 96 / 692
Galaxy S26 21:41 encoder, vocoder, conditioning: CPU (XNNPACK); decoder: NPU (Qualcomm HTP) (first run: compiles and writes the NPU cache) 4 _npu_fp32 decoder; the rest this repository's 0.0556 0.0480 2.309 B3-B5 pass; B1, B2 to come B5 min 0.990044 (0); 0 lengths differ; flips 0 / 508
Galaxy S26 21:42 encoder, decoder, conditioning: CPU (XNNPACK); vocoder: NPU (Qualcomm HTP) (first run: compiles and writes the NPU cache) 4 _npu_fp32 vocoder_frame; the rest this repository's 0.0693 0.0708 5.166 B3-B5 pass; B1, B2 to come B5 min 0.993620 (0); 0 lengths differ; flips 0 / 0
Galaxy S26 21:42 encoder, decoder, conditioning: CPU (XNNPACK); vocoder: NPU (Qualcomm HTP) (with the NPU cache of the run before) 4 _npu_fp32 vocoder_frame; the rest this repository's 0.0684 0.0678 3.215 FAIL (B3-B5) B5 min 0.988647 (1); 0 lengths differ; flips 0 / 0
Pixel 11 Pro 21:16 CPU (XNNPACK), every graph 4 this repository's 0.0332 0.0341 2.917 PASS B1 0 transcripts differ; B2 min 0.999998 (0); B5 min 1.000000 (0); 0 lengths differ; flips 0 / 0
Pixel 11 Pro 21:17 CPU (XNNPACK), every graph 1 this repository's 0.0395 0.0387 2.199 PASS B1 0 transcripts differ; B2 min 0.999998 (0); B5 min 1.000000 (0); 0 lengths differ; flips 0 / 0
Pixel 11 Pro 21:18 CPU (XNNPACK), every graph 4 this repository's 0.0356 0.0393 2.817 PASS B1 0 transcripts differ; B2 min 0.999998 (0); B5 min 1.000000 (0); 0 lengths differ; flips 0 / 0
Pixel 11 Pro 21:52 CPU (XNNPACK), every graph 4 _wfp16 conditioning, decoder, encoder, vocoder_frame 0.0340 0.0366 2.888 FAIL (B3-B5) B5 min 0.988352 (1); 5 lengths differ; flips 6 / 188

Still to run when this page was generated: 12 configurations on the Pixel 11 Pro.

Samples

“NASA is ready for the next mission.”, the author's sample sentence, in both voices.

File Voice Made by Samples
samples/pi_litert.wav Pi runner/pivo_runner.cc (LiteRT 2.3.0 C API) on a Galaxy S26 CPU (1 thread), this repository's fp32 files, vocoder_frame, 2026-11-08; its float32 audio written as PCM16 with the author's wav_bytes 29920
samples/pi_reference.wav Pi examples/say.py --check: the author's code (host/pivo/) on ONNX Runtime 1.31.1, CPU, intra-op 1, the author's ONNX files at 515d84fc 29920
samples/vo_litert.wav Vo runner/pivo_runner.cc (LiteRT 2.3.0 C API) on a Galaxy S26 CPU (1 thread), this repository's fp32 files, vocoder_frame, 2026-11-08; its float32 audio written as PCM16 with the author's wav_bytes 28960
samples/vo_reference.wav Vo examples/say.py --check: the author's code (host/pivo/) on ONNX Runtime 1.31.1, CPU, intra-op 1, the author's ONNX files at 515d84fc 28960
samples/spectrogram.png Pi, Vo log-mel power of the four WAV files (host/pivo_check.py), matplotlib -
  • Pi: LiteRT 29,920 samples, reference 29,920; log-mel correlation 0.9999993, HNR difference -0.0000 dB, 2-8 kHz share difference +0.0000 dB (host/pivo_check.py).

  • Vo: LiteRT 28,960 samples, reference 28,960; log-mel correlation 0.9999997, HNR difference -0.0000 dB, 2-8 kHz share difference -0.0000 dB (host/pivo_check.py).

  • The Pi reference file against the author's own samples/pi.wav: the same length, 343 samples differ, by at most 1 (PCM16 units).

  • The Vo reference file against the author's own samples/vo.wav: the same length, 407 samples differ, by at most 1 (PCM16 units).

Scripts

Script What it does
examples/say.py Turns text into a WAV with the LiteRT graphs. --check adds the reference audio and the B3-B5 and length comparison
host/pivo_litert.py The Python host: the author's frontend and synthesis loop, running the four LiteRT graphs through CompiledModel (CPU, or GPU at FP32)
host/pivo_check.py Computes B3 (HNR), B4 (2-8 kHz share), B5 (log-mel correlation, with DTW) and the length check
conversion/build_graphs.py, pivo_lt.py, lt_tools.py Build the graphs from the author's ONNX files: fp32, plus the float16-weight, int8 and NPU forms
conversion/check_graphs.py Line A: each graph against the author's ONNX graph on the same inputs
runner/pivo_runner.cc, build.sh, fetch_kit.sh The C++ runner (LiteRT C API) behind the phone results, its build script for Android arm64 and macOS, and the script that fetches the LiteRT 2.3.0 kit
runner/make_inputs.py, runner/check_host_tables.py Write the runner's phone-id inputs, and check its RoPE and phase tables against the Python host

B1 and B2 run Whisper and ECAPA in PyTorch. That scorer is not in this repository; the versions it used are listed under "Scorers".