shockz1/parakeet-tdt_ctc-110m-fp16-onnx

Parakeet TDT-CTC 110M — ONNX (fp16)

Half-precision ONNX export of NVIDIA's parakeet-tdt_ctc-110m, packaged for in-browser speech recognition with onnxruntime-web on WebGPU.

Used by the subtitle tool in noupload — the weights are fetched once, cached in the browser, and every subsequent run is fully offline. No audio is ever uploaded anywhere.

Files

Single-file exports: no external .onnx_data, so a browser can load each one with a single fetch.

Usage

import { ParakeetModel } from "parakeet.js";

const base = "https://huggingface.co/shockz1/parakeet-tdt_ctc-110m-fp16-onnx/resolve/main";
const model = await ParakeetModel.fromUrls({
  encoderUrl: `${base}/encoder-model.fp16.onnx`,
  decoderUrl: `${base}/decoder_joint-model.fp16.onnx`,
  tokenizerUrl: `${base}/vocab.txt`,
  backend: "webgpu-hybrid",
  nMels: 80, // this model uses 80 mel bins, not the 128 of the 0.6B models
});

const { words } = await model.transcribe(pcm16kMono, 16000, { returnTimestamps: true });

Note: parakeet.js sizes the prediction network for the 0.6B models (2 layers). This one has 1 layer, so read the real shape off joinerSession.inputMetadata for input_states_1 and resize the initial decoder state accordingly.

Provenance

Converted to fp16 from the fp32 ONNX export in `OpenVoiceOS/nvidia-parakeet-tdt_ctc-110m-onnx`, itself an export of `nvidia/parakeet-tdt_ctc-110m`. Licensed CC-BY-4.0, following the upstream model.