litert-community/Speaker-Diarization-LiteRT

Measured on device (edge-compat): Galaxy S26 · LiteRT 2.2.0 · GPU (ML Drift) · 6.05 ms p50 (2026-08-25); Galaxy S26 · LiteRT 2.2.0 · NPU (QNN/HTP) · 1.99 ms p50 (2026-08-25); Raspberry Pi 5 · LiteRT 2.2.0.dev20260804 · CPU/XNNPACK, 4 threads · 186 ms p50 (2026-08-31); browser · Chromium 151 on M4 Max · LiteRT.js 2.5.3 · WebGPU · 3.85 ms p50 · output matches CPU (2026-08-11). Record: https://github.com/john-rocky/edge-compat/blob/main/cards/speaker-diarization/CARD.md

Speaker Diarization — LiteRT on-device stack (pyannote 3.1 recipe)

Who-spoke-when timeline from the on-device pipeline (2-speaker conversation). Colored bars = per-speaker turns.

On-device speaker diarization ("who spoke when") for Android, following the pyannote/speaker-diarization-3.1 recipe (MIT):

file	model	runtime	license
wespeaker_emb_fp16.tflite	WeSpeaker ResNet34 speaker embedding (6.6 M)	LiteRT CompiledModel GPU	CC-BY-4.0 (weights)
pyannote_seg30.onnx	pyannote segmentation-3.0 (PyanNet SincNet+BiLSTM, 1.5 M)	onnxruntime CPU	MIT

The segmentation BiLSTM has no mobile-GPU kernel, so it runs on CPU (tiny and fast); the heavy embedding CNN runs fully on the GPU. Verified on a Pixel 8a (Tensor G3): embedding 108 / 108 nodes LITERT_CL (full residency, 1 partition), ~1.2 ms per window, device-vs-PyTorch cosine 0.99997 (fp16, 13.4 MB); segmentation ONNX corr 1.0 / per-frame argmax agreement 100% vs PyTorch.

I/O

Embedding wespeaker_emb_fp16.tflite

• Input [1, 500, 80] float32 — kaldi log-mel fbank (25 ms / 10 ms hamming, 80 bins, dither 0, waveform ×2¹⁵ before fbank), CMN'd (subtract the per-bin mean over the 500 frames). 500 frames = 80 240 samples = 5.015 s @ 16 kHz; tile-pad shorter speech.
• Output [1, 256] — speaker embedding (L2-normalize before cosine comparison).

Segmentation pyannote_seg30.onnx

• Input [1, 1, 160000] float32 — 10 s @ 16 kHz mono, [-1, 1].
• Output [1, 589, 7] — per-frame log-probs over the powerset classes {∅, s1, s2, s3, s1s2, s1s3, s2s3} (≤3 local speakers, ≤2 concurrent).

Minimal usage (Python)

import numpy as np, soundfile as sf, torch, onnxruntime as ort
import torchaudio.compliance.kaldi as kaldi
from ai_edge_litert.interpreter import Interpreter

wav, sr = sf.read("speech.wav", dtype="float32")     # 16 kHz mono, [-1, 1]

# 1) segmentation: who is active in a 10 s window
seg = ort.InferenceSession("pyannote_seg30.onnx")
x = np.zeros(160000, np.float32); n = min(len(wav), 160000); x[:n] = wav[:n]
ps = seg.run(None, {"waveform": x[None, None]})[0][0]          # [589, 7] log-probs
PS = [(), (0,), (1,), (2,), (0, 1), (0, 2), (1, 2)]            # powerset classes
active = [PS[c] for c in ps.argmax(1)]                          # local speakers per ~17 ms frame

# 2) speaker embedding of a 5.015 s snippet (tile-pad shorter speech to 80240 samples)
snip = np.resize(wav, 80240).astype(np.float32)
fb = kaldi.fbank(torch.tensor(snip[None]) * 32768, num_mel_bins=80, frame_length=25.0,
                 frame_shift=10.0, dither=0.0, window_type="hamming", sample_frequency=16000)
fb = (fb - fb.mean(0, keepdim=True))[None].numpy()              # CMN -> [1, 500, 80]
emb = Interpreter(model_path="wespeaker_emb_fp16.tflite"); emb.allocate_tensors()
emb.set_tensor(emb.get_input_details()[0]["index"], fb); emb.invoke()
e = emb.get_tensor(emb.get_output_details()[0]["index"])[0]     # [256]
e /= np.linalg.norm(e)                                          # cosine-compare across snippets,
                                                                # cluster at distance 0.7046

Kotlin (Android)

// Embedding — LiteRT CompiledModel GPU:  implementation("com.google.ai.edge.litert:litert:2.1.5")
val emb = CompiledModel.create(File(ctx.filesDir, "wespeaker_emb_fp16.tflite").absolutePath,
    CompiledModel.Options(Accelerator.GPU), null)
val eIn = emb.createInputBuffers()
val eOut = emb.createOutputBuffers()
eIn[0].writeFloat(fbankCmn)            // [500 * 80]: kaldi fbank (25/10 ms hamming, x2^15) + CMN,
emb.run(eIn, eOut)                     //   see Fbank.kt in the speaker_diarization LiteRT sample
val e = eOut[0].readFloat()            // [256] — L2-normalize, cosine-compare, cluster at 0.7046

// Segmentation — onnxruntime CPU:  implementation("com.microsoft.onnxruntime:onnxruntime-android:1.24.3")
val env = OrtEnvironment.getEnvironment()
val seg = env.createSession(File(ctx.filesDir, "pyannote_seg30.onnx").absolutePath,
    OrtSession.SessionOptions())
OnnxTensor.createTensor(env, FloatBuffer.wrap(window), longArrayOf(1, 1, 160000)).use { t ->
    seg.run(mapOf(seg.inputNames.first() to t)).use { out ->
        @Suppress("UNCHECKED_CAST")
        val ps = (out[0].value as Array<Array<FloatArray>>)[0]   // [589][7] powerset log-probs
        // per-frame argmax -> {∅, s1, s2, s3, s1s2, s1s3, s2s3}
    }
}

Pipeline (as in the reference)

Sliding 10 s windows → powerset argmax → per-(window, local speaker) units with enough solo speech → embedding of each unit's concatenated solo audio → agglomerative clustering (centroid linkage, cosine distance, threshold 0.7046 from the 3.1 config) → stitched global timeline.

Conversion

Embedding converted with litert-torch from pyannote/wespeaker-voxceleb-resnet34-LM: a pure CNN (no maxpool stem) — zero re-authoring except the StatsPool standard deviation (down-scaled unbiased variance, fp16-safe). fp16 tflite vs PyTorch cosine 1.0000. Segmentation exported to ONNX from pyannote/segmentation-3.0.

Upstream

• pyannote.audio (MIT) — please cite Bredin 2023 (pyannote 2.x/3.x) when you use these models.
• WeSpeaker (Apache-2.0 code; the voxceleb-resnet34-LM weights are CC-BY-4.0).

Performance

Measured on a Pixel 8a (Tensor G3, Android 16) with the standard TFLite benchmark_model tool — 10 warm-up runs then 50 timed runs, reported as the tool's mean.

Runtime	Backend	Graph on GPU	Latency
LiteRT CompiledModel (LITERT_CL)	GPU	108 / 108	~1.2 ms
TFLite benchmark_model (TfLiteGpuDelegateV2)	GPU (OpenCL)	108 / 108	21.5 ms
TFLite benchmark_model	CPU (XNNPACK, 4 threads)	—	XNNPACK declined the graph

The two GPU rows are different runtimes, not a contradiction. The LITERT_CL figure is the one recorded when this model shipped, taken through LiteRT's own CompiledModel accelerator — the path the Kotlin sample app and the LiteRT API use. The TfLiteGpuDelegateV2 figure is the classic TFLite OpenCL delegate, measured with a tool anyone can download and re-run. They agree on how much of the graph the GPU takes; they disagree on speed, and the classic delegate is the slower of the two here. Read the TfLiteGpuDelegateV2 row as a reproducible floor, not as this model's speed on LiteRT.

XNNPACK declines these fp16 graphs — it reports failed to delegate DEPTHWISE_CONV_2D and then fails to allocate tensors — so there is no usable CPU number. Disabling XNNPACK falls back to reference kernels, which measured about 20× slower than the GPU on models of this size and would not represent CPU inference anyone would ship.

Snapdragon NPU (Hexagon)

The NPU is 3.04x faster than the GPU (1.99 ms against 6.05 ms) and loads 6.81x faster (103 ms against 703 ms).

backend	compiled	inference (median / min)	load
NPU (Hexagon v81)	on-device JIT	1.99 ms / 1.95 ms	103 ms
GPU (Adreno)	—	6.05 ms / 5.76 ms	703 ms

Measured on a Samsung Galaxy S26 (Snapdragon 8 Elite Gen 5 / SM8850, Hexagon v81, Android 16) with LiteRT CompiledModel 2.2.0, one accelerator per process, 5 warm-up runs then N=50 timed runs, median reported. Every run held thermal status NONE throughout. Headroom 0.80, where 1.0 is the throttling threshold.

The NPU rows ran the published file unchanged. LiteRT compiled it for the Hexagon on the device at first load. That first compile took 630 ms here. The load column above is the cached load every later run pays. Recipe and the runtime libraries it needs: NPU guide.

GPU wiring: GPU guide.

Raspberry Pi 5 (CPU)

Measured on a Raspberry Pi 5 Model B Rev 1.1 (8 GB, Raspberry Pi OS 64-bit) with the LiteRT benchmark_model tool from litert-cli-nightly 0.2.0.dev20260805: CPU inference (XNNPACK, 4 threads), 3 invocations per file of 10 warm-up plus 50 timed runs (the tool caps a phase at 150 s, so very slow graphs run fewer — the Runs column is the actual timed total). The latency is the median across invocations; the spread is the min–max over all timed runs. No thermal throttling occurred during these runs (vcgencmd get_throttled stayed 0x0).

File	Inference (median)	Spread (min–max)	Runs	Peak memory
wespeaker_emb_fp16.tflite	186.1 ms	185.5–191.6 ms	150	138 MB