@osanseviero could you take a look at my amazing Spanish skills here? haha

Yes, your Spanish was flawless 🚀 nice!

DeepL really is not bad then 😅

Can we select the four first indices then still pass a dataset? Using a list like this contradicts the text above (you can pass a whole dataset to a pileline).

I think the blocker for this is that the pipeline expects

• A raw ndarray (here d["array"])
• A dict of {"raw": ndarray, "sampling_rate": 16_000} but datasets creates {"array": ndarray, "sampling_rate": 16_000}

speech_recognizer(dataset["audio"]["array"])
....

Should work no ? (since the sampling rate is already taken care of beforehand)

dataset["audio"] loads + resamples the whole dataset. For audio it's important to first slice the list and then call the columns.

We could use dataset.select(...) to remedy this but it'd still not work, e.g.:

from datasets import load_dataset

ds = load_dataset("PolyAI/minds14", "en-US", split="train")
ds = ds.select(range(4))

ds["audio"]["array"] # <- gives TypeError: list indices must be integers or slices, not str

Think the best I can do here is to rewrite the text a bit, no?

cc @lhoestq @albertovilla @polinaeterna @anton-l here as well - think that's a common problem we have with ds["audio"]["array"]

Ok for me to rewrite the text and say it can take list/array/dataset. But the previous example looked simpler :-p

Agree - @lhoestq @albertovilla do you have a nifty trick here by any chance to allow one to pass ds[:4]["audio"]["array"] ?

You can't do ds[:4]["audio"]["array"]. Indeed in the general case, the sampling rate may not be the same for all samples: sampling_rate is an optional parameter of Audio. So the format of ds[:4]["audio"] is a list of dicts {"array":..., "sampling_rate":..., "path":...}

We can update the pipeline to accept both raw and array. Seems like the path of least resistance here, wdyt ?

Changing the text and merging for now though

Same here.