soniqo/Kokoro-82M-LiteRT

Kokoro 82M — LiteRT

Kokoro-82M exported as a staged FP32 LiteRT bundle for CPU text-to-speech. It produces 24 kHz Float32 audio with the included af_heart voice.

This repository contains one deployment choice: the validated three-stage, 60-frame CPU/XNNPACK bundle.

Model

Property	Value
Parameters	82M
Architecture	Non-autoregressive, single pass
Format	LiteRT / TFLite
Precision	FP32
Model files	333,529,644 bytes total (318 MiB)
Sample rate	24 kHz
Public text input	128 token slots
Active text encoder/recurrent window	32 token slots
Output tensor	60 frames / 36,000 samples / 1.5 seconds
Runtime safety boundary	56 frames / 33,600 samples / 1.4 seconds
Included voice	af_heart

The recurrent stage preserves the accepted 128-slot result while evaluating only 32 slots. The backward-LSTM state generated by the guaranteed 96-slot zero tail is precomputed, and the remaining input projection is performed once per sequence outside compact LiteRT loops.

Longer text is supported by host-side chunking. The model-reported duration is checked after every invocation. Output above 56 frames must be discarded and retried as smaller text; the final four tensor frames are reserved as right context.

Files

File	Bytes	SHA-256
kokoro-encoder.tflite	47,877,656	60ac2065c78cc5ec724b3a43beb6789c084a03121018f664c561c131ce65ecaa
kokoro-recurrent-equivalent32.tflite	45,004,644	560b6318be2ac91dbeba928a634097fc6f1c30229b016f9dedacc0a6a64d2865
kokoro-vocoder.tflite	240,647,344	d8cc32713beeab8fadfb4190d5a7ada4d832ae2ecd16b82570c891f02d0e6ff6
config.json	<2 KB	Runtime geometry and artifact hashes
vocab_index.json	2,501	IPA symbol-to-token mapping
us_gold.json	3,000,469	Primary English pronunciation dictionary
us_silver.json	3,099,517	Fallback English pronunciation dictionary
voices/af_heart.bin	1,024	Default 256-value Float32 voice embedding

Validation

The complete chained bundle was compared with the FP32 PyTorch source using identical token, voice, speed, and harmonic-phase inputs.

Check	Result
Waveform correlation	0.999745
Magnitude-spectrogram correlation	0.999947
Duration maximum absolute error	0
Duration mismatch count	0
Recurrent equivalence matrix	0 mismatches across 42 length/speed cases
Validated host cutoff	56 frames

The physical Android and Windows wrapper outputs correlate at 0.999795 (34.4 dB SNR). The staged Android output correlates at 0.999834 with the previous accepted monolithic Android output.

Performance

Physical-device synthesis through the speech-core C++ wrapper on a Samsung Galaxy S23 Ultra (SM-S918B, Snapdragon 8 Gen 2), Android 16, LiteRT 2.1.5, XNNPACK, all eight CPU cores (taskset ff):

Text	Load	Cold	Warm p50	Warm p90	Audio	p50 RTF	p90 RTF	Peak RSS
Hello world.	489.2 ms	465.3 ms	522.7 ms	551.7 ms	0.95 s	0.550	0.581	1,017.5 MiB

One warm-up and ten measured runs were used; model loading was excluded from RTF. Thermal status remained 0. Eight threads was faster than 4, 6, and 7 in isolated physical-device runs.

Usage with speech-core

#include <speech_core/models/litert_kokoro_tts.h>

speech_core::LiteRTKokoroTts tts(
    "kokoro-encoder.tflite",
    "kokoro-recurrent-equivalent32.tflite",
    "kokoro-vocoder.tflite",
    "voices",
    ".",
    /*hw_accel=*/false,
    /*num_threads=*/8);

tts.set_speed(1.0f);
tts.set_seed(1234);
tts.synthesize("Hello world.", "en",
    [](const float* pcm, size_t samples, bool is_final) {
        // 24 kHz mono Float32 PCM
    });

The wrapper includes the English phonemizer, proactive token chunking, the 56-frame duration guard, retry splitting, waveform checks, and final audio assembly.

Build and run the physical benchmark:

cmake --build speech-core/build --target speech_kokoro_litert_bench
speech_kokoro_litert_bench /path/to/this/bundle \
  --variant equivalent32 --threads 8 --warmup 1 --runs 10

Source

Converted from hexgrad/Kokoro-82M. The bundle uses explicit harmonic phases, compact FP32 recurrent loops, and a LiteRT-compatible iSTFT path.