V3 audio transcription: aud.subarray is not a function

[flatsiedatsie]

System Info

Cutting edge version of V3 (just compiled)

Environment/Platform

• Website/web-app
• Browser extension
• Server-side (e.g., Node.js, Deno, Bun)
• Desktop app (e.g., Electron)
• Other (e.g., VSCode extension)

Description

I attempted a drop-in replacement of the V3 version in a V2 webworker, just to see if it would even work.

And it does! Trying to enable WebGPU is the next step.

However, while it works, I do see an error:

Reproduction

I could share code if need be.

[flatsiedatsie]

I spotted a small typo in the demo:

PipelineSingeton is missing an L

[flatsiedatsie]

A small question: is there a downside to simply always using the timestamp version?

[flatsiedatsie]

And some more questions:

The V2 demo has a number of values that could be manipulated. Are they still useful?

• const isDistilWhisper = model.startsWith("distil-whisper/");
• quantized

and these:

	        /// Greedy
	        //top_k: 0,
	        //do_sample: false,

	        // Sliding window
	        //chunk_length_s: isDistilWhisper ? 20 : 30,
	        //stride_length_s: isDistilWhisper ? 3 : 5,

	        // Language and task
	        //language: language,
	        task: subtask,

	        // Return timestamps
	        //return_timestamps: true,
	        //force_full_sequences: false,

[flatsiedatsie]

I saw this error once:

[flatsiedatsie]

I'm having some trouble getting it to respond with more than 1 word when using WebGPU. It usually thinks it heard 'And'.

[flatsiedatsie]

FP32 was the key

[xenova]

Thanks for the report! The function does require Float32Array or Float64Array inputs, but we could use .slice() if .subarray isn't present (for normal arrays)

[flatsiedatsie]

Nah, no worries. With what you're saying I believe the issue was that I was feeding it a fake array to get it to preload. That worked with the V2 version, but no longer worked with the V3 version. But that's fine as there are plenty of new ways to handle pre-loading.