BuilderHelm/Parakeet-CTC-110M-CoreML

Parakeet-CTC-110M CoreML

NVIDIA's Parakeet-TDT-CTC-110M model converted to CoreML for fast CTC transcription and keyword spotting on Apple Silicon, by BuilderHelm AI.

Need TDT transcription too? Use BuilderHelm/Parakeet-TDT-CTC-110M-CoreML, which has the same encoder plus the TDT decoder.

Highlights

• Keyword spotting: 89.8% recall at one false alarm per clip on Earnings22
• Accurate CTC: within 0.1 points of the full-precision model on LibriSpeech
• Neural Engine: int8 encoder runs on the ANE
• Fast: 1,005× real time, 11.5 ms per dictation clip (M5)
• Any length: 4, 8, 15 and 20-second windows; longer audio is split automatically

Model Description

The CTC path of NVIDIA's hybrid Parakeet-TDT-CTC-110M: a shared FastConformer encoder and its CTC head. It gives a greedy transcript and per-frame log-probabilities for keyword spotting, custom vocabulary and alignment.

Architecture

Component	Description	Size
Preprocessor	Audio → log-mel (float32, CPU)	5.7 MB
Encoder	FastConformer + CTC head (int8, Neural Engine)	112.7 MB
vocab.json	1,024-token vocabulary	9 KB
config.json	Model configuration and I/O contracts	4 KB

Total size: 118 MB

Input: 16 kHz mono audio, any length. Output: CTC transcript and log-probabilities every 80 ms (1,025 classes; blank = 1024).

Performance

Apple M5, macOS 27.2. Same audio and scoring for every system (full evaluation).

CTC (WER %):

System	Common Voice	Earnings22	LS clean	LS other	Long audio
This model	8.29	17.62	2.97	5.88	3.19
Full-precision reference	8.65	17.59	2.94	5.92	2.77
FluidInference CTC 110M	94.86	99.27	97.27	98.51	96.93
mudler q8_0 GGUF	8.65	17.53	2.95	5.92	2.76

Speed (× real time):

System	Common Voice	Earnings22	LS clean	LS other	Long audio
This model	1,005	1,235	1,003	995	991
FluidInference CTC 110M	95	146	78	73	130
mudler q8_0 GGUF	305	345	280	279	86

Custom Vocabulary / Keyword Spotting

Run CTC word spotting over ctcLogProbs with your terms tokenized by vocab.json. Keyword recall on Earnings22 (772 clips, about 25 candidate terms per clip), same spotter for both models:

False alarms per clip	This model	FluidInference CTC 110M
0.1	46.3%	1.1%
0.5	85.4%	2.8%
1.0	89.8%	5.0%

With only the spoken terms as candidates, recall is 97.6% (this model) and 99.2% (FluidInference).

Requirements

• macOS 15+ on Apple Silicon
• Swift 6 toolchain (for the Swift package)

Installation

pip install -U huggingface_hub

hf download BuilderHelm/Parakeet-CTC-110M-CoreML --revision v1.0.0 \
  --include "*.mlpackage/*" --include "config.json" --include "vocab.json" --include "swift/*" \
  --local-dir parakeet-ctc-110m

Or download the archive (105 MB).

Usage

Swift

Add the swift/ folder as a local Swift Package:

import Foundation
import ParakeetTdtCtc110M

let model = try ParakeetTdtCtc110M(directory: URL(fileURLWithPath: "parakeet-ctc-110m"), ctcOnly: true)
let samples = try ParakeetAudio.read(URL(fileURLWithPath: "audio.wav"))
let result = try model.transcribe(samples)

print(result.text)                  // CTC transcript
let logProbs = result.ctcLogProbs   // [frame][1025], blank = 1024

Command Line

cd parakeet-ctc-110m
swift run -c release --package-path swift parakeet-110m . audio.wav --ctc

File Structure

./
├── Preprocessor.mlpackage    # Audio → log-mel
├── Encoder.mlpackage         # Log-mel → CTC log-probs
├── vocab.json                # Token vocabulary (1,024 tokens)
├── config.json               # Model configuration
├── compiled/                 # Precompiled .mlmodelc (macOS 27.2)
├── releases/                 # ZIP archive
└── swift/                    # Swift library and CLI

Limitations

• English only
• First load takes about 18 seconds while CoreML compiles the models
• Tested on Apple M5 (macOS 27.2) only
• Settings were tuned on parts of the test sets

Model Conversion

Built from NVIDIA's checkpoint with the recipe in Parakeet-TDT-CTC-110M-CoreML. These files are byte-identical to the Preprocessor and Encoder there.

License

Model weights and tokenizer: CC-BY-4.0, from NVIDIA. Code: MIT. See NOTICE.md.

Citation

@misc{nvidia_parakeet_tdt_ctc_110m,
  title  = {Parakeet-TDT-CTC-110M},
  author = {NVIDIA},
  year   = {2024},
  url    = {https://huggingface.co/nvidia/parakeet-tdt_ctc-110m}
}

Acknowledgments

• Original model by NVIDIA NeMo
• CoreML conversion by BuilderHelm AI