mattt/parakeet-ultra-coreml

Parakeet Ultra for Core ML

Core ML components of moondream/parakeet-ultra for AnySpeechRecognizer.

Attribution and license

Parakeet Ultra by Moondream (https://huggingface.co/moondream/parakeet-ultra), a post-trained version of Parakeet TDT 0.6B v3 by NVIDIA (https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3); both licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Converted to Core ML for AnySpeechRecognizer: the model is split into encoder, prediction network, and joint decision components, weights are stored in float16, and encoder weights are quantized to int8.

This repository is licensed under CC BY 4.0, the license of the base model. Redistributions must keep this attribution.

Files

File	Contents
Encoder.mlmodelc	FastConformer encoder for one 15-second window; int8 per-channel weights.
Encoder8s.mlmodelc	FastConformer encoder for one 8-second window (801 mel frames, 101 encoder frames); int8 per-channel weights.
Decoder.mlmodelc	Prediction network (two-layer LSTM), one step with explicit state.
JointDecisionv3.mlmodelc	Joint network and greedy TDT decision for one encoder frame, with the top 64 tokens.
parakeet_vocab.json	Token ID to SentencePiece piece, with the word-boundary marker written as a space.
conversion.json	Sources, tool versions, settings, and SHA-256 digests of every file.

The models take NeMo-compatible 128-bin log-mel features for a 15-second window (1,501 frames) and produce 188 encoder frames of 80 ms each. The tensor interface matches FluidInference's Encoder_v2.mlmodelc, Decoder.mlmodelc, and JointDecisionv3.mlmodelc artifacts. The preprocessor is not included.

Each Encoder<seconds>s.mlmodelc is the same encoder traced for a shorter fixed window, with the same trained weights, input and output names, and 80 ms encoder frames. It shares the decoder and joint decision.

Conversion

• Sources:
  • nvidia/parakeet-tdt-0.6b-v3 at revision 541d1f99c6b0c3cd0b11a95167540bb8edefd82b: parakeet-tdt-0.6b-v3.nemo (SHA-256 3cbdc85877e668ca7b82d0d56770eb1fac76691f55d6b97545e8d61ca588d10d), architecture, tokenizer, preprocessor.
  • moondream/parakeet-ultra at revision 73175eb7aeb0d82f1e2a6b53b3aabc10a90bcd0b: model.safetensors (SHA-256 c9608f36d0ab956c14bfcc525479b0746b3b42a56f6949ec85c14eb7466717dc), weights.
  • moondream/parakeet-ultra at revision 73175eb7aeb0d82f1e2a6b53b3aabc10a90bcd0b: tokenizer.json (SHA-256 bd321b096832a3f270bd3b2a88823957920f1a5c5ada71114a26ea729d0cbe91), tokenizer check.
• Conversion: TorchScript trace to an ML program with an iOS 17 and macOS 14 deployment target and float16 compute precision.
• Encoder weights: int8 linear symmetric quantization, one scale per output channel, for weights with at least 2,048 elements.
• Tools: Python 3.12.12, PyTorch 2.7.0, NeMo 2.3.1, coremltools 9.0, coremlcompiler 3600.25.1, macOS 27.0.1.
• The conversion follows FluidInference/mobius (Apache-2.0) at commit 26f5fa8059d5ab3d34449adce6ff6ed86734e522.