[Bug]: TTS fails to parse 429 warning message

[kyu-aurelian]

Summary

when DG returns 429 error message, it fails to parse and completely fails the request

What happened?

it should just call on_warning call back, but it's entirely failing the connection.

Steps to reproduce

n/a

Minimal code sample

n/a

Logs / traceback

I think there is a gap in how DSK handles warning message between the old and new SDK version.
message:

{'type': 'Warning', 'warn_code': 'EXCESSIVE_FLUSH', 'warn_msg': 'Rate limit exceeded for flushes. Please try again later.'}

stack trace from the new SDK:

Traceback (most recent call last):
File "/app/call_handling/consumers/synthesizer/provider/deepgram.py", line 140, in _listen
await connection.start_listening()
File "/app/.venv/lib/python3.11/site-packages/deepgram/speak/v1/socket_client.py", line 84, in start_listening
parsed, is_binary = self._process_message(raw_message)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/app/.venv/lib/python3.11/site-packages/deepgram/s\peak/v1/socket_client.py", line 62, in _process_message
processed = self._handle_json_message(raw_message)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/app/.venv/lib/python3.11/site-packages/deepgram/speak/v1/socket_client.py", line 54, in _handle_json_message
return parse_obj_as(V1SocketClientResponse, json_data) # type: ignore
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/app/.venv/lib/python3.11/site-packages/deepgram/core/pydantic_utilities.py", line 43, in parse_obj_as
return adapter.validate_python(dealiased_object)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/app/.venv/lib/python3.11/site-packages/pydantic/type_adapter.py", line 421, in validate_python
return self.validator.validate_python(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
pydantic_core._pydantic_core.ValidationError: 10 validation errors for union[bytes,SpeakV1MetadataEvent,SpeakV1ControlEvent,SpeakV1WarningEvent]
bytes
Input should be a valid bytes [type=bytes_type, input_value={'type': 'Warning', 'warn...lease try again later.'}, input_type=dict]

async def start_listening(self):
  ...
  try:
    async for raw_message in self._websocket:
      parsed, is_binary = self._process_message(raw_message)
  ...

def _handle_json_message(self, message: str) -> typing.Any:
  """Handle a JSON message by parsing it."""
  json_data = json.loads(message)
  return parse_obj_as(V1SocketClientResponse, json_data)  # type: ignore

In the new SDK, it just tries to parse it into V1SocketClientResponse (SpeakV1WarningEvent specifically), but there is mismatch between property names, and it fails:

class SpeakV1WarningEvent(UniversalBaseModel):
    """
    Warning event from the TTS WebSocket
    """
    
    type: typing.Literal["Warning"]
    """Message type identifier"""
    
    description: str
    """A description of what went wrong"""
    
    code: str
    """Error code identifying the type of error"""

SpeakV1WarningEvent.description
Field required [type=missing, input_value={'type': 'Warning', 'warn...lease try again later.'}, input_type=dict]
For further information visit https://errors.pydantic.dev/2.11/v/missing
SpeakV1WarningEvent.code
Field required [type=missing, input_value={'type': 'Warning', 'warn...lease try again later.'}, input_type=dict]
For further information visit https://errors.pydantic.dev/2.11/v/missing

But in the previous SDK, it just handled this warning message correctly/gracefully.
previous SDK:

async def _process_text(self, message: Union[str, bytes]) -> None:
  ...
  data = json.loads(message)
  response_type = data.get("type")
  ...
  match response_type:
    ...
    case SpeakWebSocketEvents.Warning:
      war_warning: WarningResponse = WarningResponse.from_json(message)
      self._logger.verbose("WarningResponse: %s", war_warning)
      await self._emit(
        SpeakWebSocketEvents(SpeakWebSocketEvents.Warning),
        warning=war_warning,
        **dict(cast(Dict[Any, Any], self._kwargs)),
      )
    ...

@dataclass
class WarningResponse(BaseResponse):
    """
    Warning Message from the Deepgram Platform
    """

    warn_code: str = ""
    warn_msg: str = ""
    type: str = ""

In addition to this property name discrepancy, I see previous SDK uses cls.from_dict, which I think won't hard fail even property names mismatch, while the new SDK uses pydantic, which I think is more strict about this.

Transport

HTTP

API endpoint / path

speak.v1.connect

Model(s) used

aura-2

How often?

Always

Is this a regression?

• Yes, it worked in an earlier version

Last working SDK version (if known)

4.8.1

SDK version

5.1.0

Python version

3.11

Install method

None

OS

Linux (x86_64)

Environment details

Link to minimal repro (optional)

No response

Session ID (optional)

No response

Project ID (optional)

No response

Request ID (optional)

No response

Code of Conduct

• I agree to follow this project’s Code of Conduct

[Jacob-Lasky]

Summarizing here: Aura is returning a warning message with keys like warn_code and warn_msg, but the v5.1 SDK expects code and description. Because the new SDK relies on Pydantic v2, which validates unions more strictly, that mismatch throws a ValidationError and closes the stream instead of routing the event to on_warning.