server: accept base64-encoded voice reference in speech requests

[IIIIIllllIIIIIlllll]

Add base64 support to the speech API. During use, I noticed that every time I clone a voice, I have to upload the audio file to the server first, and then use the file path for audiocpp_server to read it—which is quite inflexible. Therefore, I plan to add this feature so that clients can directly upload audio files encoded in base64.

Move base64 helpers into app/server/base64.{h,cpp}, add a strict decoder, and cover it with a server_base64_test unittest.

The maximum file size needs to be considered—I'm concerned that users might inadvertently upload extremely large audio files, causing machines with limited memory to run out of memory (OOM) directly.

What do you think?

[0xShug0]

@IIIIIllllIIIIIlllll Thanks! I like the feature. For reference voice, models usually recommands 5 to 15s. We can assume a safe max is 30s (base64 < 5MB).

I wonder if we can avoid adding a separate top-level voice_ref_b64 option by overloading the voice_ref option?

"voice_ref": {
  "type": "base64",
  "data": "UklGRh..."
}

"voice_ref": {
  "type": "path",
  "path": "voices/alice.wav"
}

"voice_ref": "voices/alice.wav"

[IIIIIllllIIIIIlllll]

I wonder if we can avoid adding a separate top-level voice_ref_b64 option by overloading the voice_ref option?

OK, I think we can distinguish between a file path and base64-encoded data by checking the beginning of the string.

Edited:
Your method is more reliable after all.

[0xShug0]

@IIIIIllllIIIIIlllll Sorry, just for clarification: "voice_ref": "voices/alice.wav" is backward-compatible, so existing users can continue using it, or they can use:

"voice_ref": {
  "type": "path",
  "path": "voices/alice.wav"
}

And for Base64, the user must explicitly send it.

"voice_ref": {
  "type": "base64",
  "data": "UklGRh..."
}

So you don't need to distinguish between them based on prefix.

PS: In fact, we can’t prevent users from sending random stuff. Comprehensive checks are better handled in the frontend or gateway before requests reach the HTTP server so the server remains lightweight.