Modelpublic

KitKatL/Hojo-ASR-Multi-V1-MLX-4bit

sourceHugging Faceapache-2.0updated 19d agoView on Hugging Face

0likes144downloads

Model Card

Hojo-ASR-Multi-V1 MLX mobile 4-bit

This is a private MLX conversion of `HojoAI/Hojo-ASR-Multi-V1` for Apple silicon and the Live Transcriber native Swift runtime.

It is not a drop-in replacement for the upstream hojo-asr Python package. The native Swift/MLX model implementation is maintained separately in Live Transcriber.

Conversion

—Source repository: HojoAI/Hojo-ASR-Multi-V1
—Source revision: aa1d0dfc369deb3b3235572fa151acd7b498481f
—Source checkpoint SHA-256: c046a5a26c142f5aded7331d36892293a025f579e342a27861e80509afaca88d
—Quantization: affine 4-bit, group size 64
—Quantized scope: text decoder and frontend linear layers
—Remaining frontend dtype: float16
—Weight size: 3,089,867,955 bytes across six safetensors shards
—Tensor count: 1,942

The exact architecture and generation metadata are recorded in config.json. .complete.json is the manifest validated by Live Transcriber before loading.

Compatibility

—macOS on Apple silicon with MLX/Metal
—Live Transcriber Hojo-ASR native runtime
—Automatic multilingual recognition
—Post-recording transcription; no word-level alignment

The public upstream inference interface does not expose forced-language, prompt, hotword, custom-dictionary, or speaker-adaptation controls. The Live Transcriber runtime therefore uses deterministic decoding with repetition penalty rather than exposing unvalidated generation controls.

Validation

The bundle was loaded directly from external storage with the native Swift/MLX runtime on Apple silicon. A 5.293-second German fixture produced:

guten morgen dies ist ein kurzer test für die lokale spracherkennung auf einem mac

Measured native test runtime was approximately 6.01 seconds. MLX reported an approximately 3.42 GB peak allocation during that test. These figures describe one local validation run and are not general performance guarantees.

The same bundle was also loaded on an iPhone 16 Pro Max in an isolated device probe, but distribution and production iOS integration remain separate gates.

License

The source model is published under the Apache License 2.0. See LICENSE. This repository contains a converted derivative and retains the upstream license notices.