Glimpse-Dictation/Parakeet-TDT-0.6B-V3-coreml

Parakeet TDT 0.6B V3: Core ML encoder and GGUF decoder

This repository contains the optional Apple Silicon package for Parakeet TDT 0.6B V3 (GGUF) in Glimpse. It requires the transcribe.cpp Parakeet Core ML adapter with decoder-only GGUF support. The application dependency integration is under development; these are not standalone transcription executables.

File	Download size	Purpose
parakeet-tdt-0.6b-v3-Q8_0-decoder.gguf	19.48 MB	Predictor, joint network, tokenizer and model metadata
parakeet-tdt-0.6b-v3-Q8_0-encoder.mlmodelc.zip	1.09 GB	Compiled Core ML encoder

The combined download is 1.11 GB, without duplicating the encoder weights in the GGUF. For CPU/GPU transcription on Windows or Mac, use Handy's full Q8_0 GGUF instead. The Core ML package requires Apple Silicon.

Provenance

The model originates from NVIDIA Parakeet TDT 0.6B V3, licensed CC BY 4.0. Both artifacts derive from Handy's parakeet-tdt-0.6b-v3-Q8_0.gguf using transcribe.cpp's scripts/convert-parakeet-gguf-to-coreml.py and scripts/extract-parakeet-decoder.py. Extraction preserves predictor/joint weights and metadata and sets stt.parakeet.decoder_only=true.

Artifact	SHA-256
Original full GGUF	5859f77944efcd8eafa23a6350731960b2b55b2203df51f319665c807d802cc7
Decoder-only GGUF	dfcf670a00df8d49474fddea707bcfc77339789f65dde115aeb79570fc813744
Encoder ZIP	7d90a75d6c9bad2f082782545adbea0df430703dc84f90404e1fe322552a563d

Runtime requirements

Extract the ZIP next to the decoder GGUF, retaining the parakeet-tdt-0.6b-v3-Q8_0-encoder.mlmodelc directory name. A native caller must pass that directory through the Core ML session option. Glimpse-Speech detects the companion beside the -decoder.gguf file.

The encoder accepts 128 mel bins and up to 1501 frames, using a length mask. It uses FP16 Core ML computation with CPU and Neural Engine allowed and GPU excluded. The predictor/joint decoder runs on CPU. Core ML may use CPU operations; this is not a guarantee of exclusive Neural Engine execution.

Glimpse splits long recordings into chunks within the encoder capacity. Native decoder-only calls beyond capacity return an error; there is no embedded ggml encoder to fall back to. Full GGUF calls retain the normal CPU/GPU fallback.

The export supports offline TDT V3, not Unified or Nemotron streaming. It does not add custom-word biasing, streaming, or speaker diarization. FP16 output can differ from ggml; no multilingual accuracy equivalence is claimed.