Add Pydantic typing

[extrange]

Hi, we are using this library extensively at our company, and internally we have added type validation with Pydantic. If this is useful, I can update the documentation as well.

[MahmoudAshraf97]

Hello, I'm not well experienced with pydantic tbh, can you explain how this is useful or helps ease future maintainability? I know there is a lot of code duplication that this reduces but is there other benefits?

[extrange]

We have a use case where we need to serialize the output tuple[Iterable[Segment], TranscriptionInfo] into JSON, and then deserialize it back into a typed Python object/model.

Namedtuples do not natively allow nested deserialization. For example, using the following JSON:

segments, info = model.transcribe(audio, word_timestamps=True)
out = [s_asdict() for s in segments]
print(json.dumps(out, indent=2))

[
  {
    "id": 1,
    "seek": 2490,
    "start": 0.0,
    "end": 6.0,
    "text": " Ladies and gentlemen, thank you for being here and for your written representations.",
    "tokens": [
      50364,
      //...
    ],
    "avg_logprob": -0.26516544380608725,
    "compression_ratio": 1.597883597883598,
    "no_speech_prob": 0.01361083984375,
    "words": null,
    "temperature": 0.0
  },
//...
]

If we try to convert this back to a list[Segment], the words is not converted:

from faster_whisper.transcribe import Segment
[Segment(**s) for s in parsed]

Segment(
    id=1,
    seek=2490,
    start=0.0,
    end=4.62,
    text=" Ladies and gentlemen, thank you for being here and for your written representations.",
    tokens=[
        50364,
# ...
    ],
    avg_logprob=-0.26516544380608725,
    compression_ratio=1.597883597883598,
    no_speech_prob=0.01361083984375,
    words=[
        [0.0, 0.32, " Ladies", 0.310546875],
        [0.32, 0.54, " and", 0.89453125],
# ...
    ],
    temperature=0.0,
)

Currently, we are recursing through the Segment and TranscriptionInfo objects to convert them to Pydantic models (which natively allow nested deserialization), but it would be nice if the library supported this out of the box.

This should also fix #667.

[MahmoudAshraf97]

I see, but why not use dataclass? it supports recursive serialization too and will not add an extra dependency or change much in the code

[extrange]

I think dataclasses are not serializable to JSON, at least not without another library:

import json
from dataclasses import dataclass


@dataclass
class Foo:
    x: str


foo = Foo(x="bar")
json.dumps(foo)

# TypeError: Object of type Foo is not JSON serializable

[extrange]

Closed in favour of #1105