YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Kokoro-82M LiteRT (Vent / speech-core)

LiteRT (.tflite) pack for on-device Kokoro TTS via speech-core KokoroTts with VENT_KOKORO_RUNTIME=litert.

Ship revision: v1.0.2-kmodel-decoder-litert-gpu — KModel LiteRT wrapper with Metal-safe Albert embeddings (no EMBEDDING_LOOKUP), same I/O contract as workband/Kokoro-82M-ONNX.

iOS GPU (Metal)

Vent requires GPU (Metal) on iOS. The pack manifest.json includes:

"litert_export": {
  "metal_gpu_supported": true,
  "embedding_lookup_op_count": 0
}

Only consume packs with metal_gpu_supported: true and embedding_lookup_op_count: 0. Older packs (e.g. v1.0.1-kmodel-decoder-litert) abort in libLiteRtMetalAccelerator during LiteRtCreateCompiledModel because Albert EMBEDDING_LOOKUP ops cannot compile on Metal.

Sign-off: physical iPhone with Vent’s vendored libLiteRt + libLiteRtMetalAccelerator (simulator optional if Metal compile succeeds).

Files

File Role
kokoro-e2e.tflite E2E model (FP32)
vocab_index.json KModel phoneme vocab
us_gold.json, us_silver.json, dict_*.json Phonemizer assets
voices/*.bin Style vectors (row 127 in app)
manifest.json Pack revision, SHA-256, litert_export metadata

Runtime contract

Same inputs/outputs as the ONNX pack: input_ids, attention_mask, ref_s, speed, random_phasesaudio, audio_length_samples, pred_dur at 24 kHz FP32.

Build & publish

Source tooling: workband/kokoro-82m-onnx.

make rebuild-ship-litert
make inspect-litert-ops   # EMBEDDING_LOOKUP must be 0
make upload-litert LITERT_TAG=v1.0.2-kmodel-decoder-litert-gpu

License

Apache-2.0 (hexgrad/Kokoro-82M weights; export tooling same as Vent).

Downloads last month
1
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support