Add support for LiteWhisper

[xenova]

LiteASR is a compression scheme for automatic speech recognition (ASR) models that leverages the low-rank properties of activation values. Our method can compress OpenAI's Whisper encoder by up to ~50%.

Supported models: https://huggingface.co/models?library=transformers.js&other=lite-whisper&sort=trending

---

Example usage:

import { pipeline } from "@huggingface/transformers";

const transcriber = await pipeline(
  "automatic-speech-recognition",
  "onnx-community/lite-whisper-large-v3-turbo-acc-ONNX",
  { dtype: { encoder_model: "fp32", decoder_model_merged: "q4" } },
);

const audio = await read_audio(
  "https://huggingface.co/datasets/Xenova/transformers.js-docs/resolve/main/jfk.wav",
  transcriber.processor.feature_extractor.config.sampling_rate,
);

const output = await transcriber(audio);
console.log(output);
// { text: ' And so, my fellow Americans, ask not what your country can do for you, ask what you can do for your country.' }

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[decoder-sh-david]

@xenova Which browser did you test this in? I'm unable to get your code sample working in Chrome. While trying to load the model, the pipeline throws an error that's just a number 4273819248.

[xenova]

@decoder-sh-david Be sure to specify device: "webgpu" to run in-browser with the current configuration.

Alternatively, you can set dtype: "q8" if you'd like to run on CPU.

The above sample code was run with Node.js (CPU)

[decoder-sh-david]

@decoder-sh-david Be sure to specify device: "webgpu" to run in-browser with the current configuration.

Alternatively, you can set dtype: "q8" if you'd like to run on CPU.

The above sample code was run with Node.js (CPU)

That was the missing piece, thank you! Do you have any guidance on which quantizations work? I recall that for some onnx versions of whisper, some just don't work

[decoder-sh-david]

Additionally, I don't suppose that this model supports word-level timestamps, does it?

[decoder-sh-david]

@xenova does this PR also support converting this model type using the script?

I'm trying to convert it to to include timestamps in the onnx export, but I keep running into the error ValueError: Unrecognized configuration class <class 'transformers_modules.efficient-speech.lite-whisper-large-v3-turbo.6697ac2a887e3256da5defc9e8472f76a2b0f16e.configuration_lite_whisper.LiteWhisperConfig'> to build an AutoTokenizer. which indicates that transformers python may not support the model?