v2 Listen: TurnInfo messages deserialized as ListenV2Connected (fields lost)

[simonepalomba]

Summary

When using client.listen.v2.connect() with flux-general-multi, all incoming messages are deserialized as ListenV2Connected regardless of the actual type field. Semantic fields on TurnInfo messages (event, transcript, end_of_turn_confidence, audio_window_start/end, words, languages) are lost.

Environment

• deepgram-sdk: 6.1.1 (also reproduced on 5.3.3)
• Python: 3.12 (Rocky Linux 9)
• Endpoint: wss://api.deepgram.com/v2/listen
• Model: flux-general-multi

Expected

TurnInfo messages from the server should be deserialized as ListenV2TurnInfo with all fields accessible.

Actual

All messages arrive as ListenV2Connected. model_dump() returns only {type, request_id, sequence_id}. The type attribute correctly contains "TurnInfo", but the discriminator seems ignored during deserialization. No errors are raised.

Repro

import asyncio, os
from deepgram import AsyncDeepgramClient
from deepgram.core.events import EventType

async def main():
    client = AsyncDeepgramClient(api_key=os.environ["DEEPGRAM_API_KEY"])
    async with client.listen.v2.connect(
        model="flux-general-multi",
        encoding="linear16",
        sample_rate="16000",
        eot_threshold="0.7",
    ) as conn:
        def on_msg(msg):
            print(f"class={type(msg).__name__} type={getattr(msg,'type','?')} "
                  f"event={getattr(msg,'event',None)} "
                  f"transcript={getattr(msg,'transcript',None)}")

        conn.on(EventType.MESSAGE, on_msg)
        task = asyncio.create_task(conn.start_listening())
        await asyncio.sleep(0.2)

        # Send ~8s of italian speech as linear16 16kHz PCM
        with open("italian_sample.raw", "rb") as f:
            audio = f.read()
        chunk = 16000 * 2 * 80 // 1000
        for i in range(0, len(audio), chunk):
            await conn.send_media(audio[i:i+chunk])
            await asyncio.sleep(0.08)
        # Trailing silence to trigger EndOfTurn
        for _ in range(25):
            await conn.send_media(b"\x00" * chunk)
            await asyncio.sleep(0.08)
        await asyncio.sleep(2)

asyncio.run(main())

Sample output

class=ListenV2Connected type=Connected event=None transcript=None
class=ListenV2Connected type=TurnInfo event=None transcript=None
class=ListenV2Connected type=TurnInfo event=None transcript=None
class=ListenV2Connected type=TurnInfo event=None transcript=None
...

Expected class for the TurnInfo rows: ListenV2TurnInfo, with populated event/transcript/end_of_turn_confidence.

Likely cause

The v2 response union ListenV2SocketClientResponse appears to be discriminated on type, but deserialization via construct_type(..., skip_validation=True) seems to silently pick ListenV2Connected (the first member?) instead of the variant matching type.

Impact

flux-general-multi is unusable via the Python SDK — no transcripts can be extracted. Downgrading to 5.3.3 does not help (same behavior).

Workaround

Temporarily bypassing the SDK with a raw websockets client against /v2/listen and parsing JSON manually works. Happy to share a minimal working example if useful for regression testing.

Thank you

Thanks for your work on the SDK. flux-general-multi for Italian is a great product — once this is fixed it'll unblock a production voicebot deployment.

[simonepalomba]

A couple of things that may help triage:

Cross-reference to #679 / #680

This looks like the same class of Fern-generated discriminator regression that was fixed on the TTS side in #679 (PR #680, shipped in 6.1.1). Same pattern — a discriminated union (SpeakV1SocketClientResponse then, ListenV2SocketClientResponse now) whose variants share a type field, deserialized through construct_type(..., skip_validation=True) which silently picks the first member instead of matching on the discriminator. The fix shape from #680 should be directly applicable here.

Minimal standalone repro without the SDK

To confirm the bug is 100% client-side, here's a ~40-line raw-websockets client that hits the same /v2/listen endpoint with flux-general-multi and parses TurnInfo manually. All semantic fields arrive populated. Happy for this to be used as a regression test once the fix lands.

# pip install websockets
import asyncio, json, os
import websockets

URL = (
    "wss://api.deepgram.com/v2/listen"
    "?model=flux-general-multi"
    "&encoding=linear16&sample_rate=16000"
    "&language_hint=it&eot_threshold=0.7"
)

async def main():
    headers = [("Authorization", f"Token {os.environ['DEEPGRAM_API_KEY']}")]
    async with websockets.connect(
        URL, additional_headers=headers,
        ping_interval=5, ping_timeout=20,
    ) as ws:
        async def recv():
            async for raw in ws:
                if isinstance(raw, bytes):
                    continue
                msg = json.loads(raw)
                if msg.get("type") == "TurnInfo":
                    print(f"TurnInfo: event={msg.get('event')} "
                          f"transcript={msg.get('transcript')!r} "
                          f"eot_conf={msg.get('end_of_turn_confidence')}")
                else:
                    print(f"{msg.get('type')}: {msg}")

        async def send():
            with open("italian_sample.raw", "rb") as f:
                audio = f.read()
            chunk = 16000 * 2 * 80 // 1000  # 80ms of linear16 @16kHz
            for i in range(0, len(audio), chunk):
                await ws.send(audio[i:i+chunk])
                await asyncio.sleep(0.08)
            # Trailing silence to trigger EndOfTurn
            for _ in range(25):
                await ws.send(b"\x00" * chunk)
                await asyncio.sleep(0.08)
            await ws.send(json.dumps({"type": "CloseStream"}))

        await asyncio.gather(recv(), send())

asyncio.run(main())

Expected output against ~8s of Italian speech followed by silence (all fields populated, proving the server-side payload is correct):

Connected: {'type': 'Connected', 'request_id': '...', 'sequence_id': 0}
TurnInfo: event=StartOfTurn transcript='' eot_conf=0.01
TurnInfo: event=Update transcript='Buongiorno' eot_conf=0.15
TurnInfo: event=Update transcript='Buongiorno come sta' eot_conf=0.42
TurnInfo: event=EndOfTurn transcript='Buongiorno come sta' eot_conf=0.87

Same audio through client.listen.v2.connect() on 6.1.1 produces the event=None transcript=None output from the original report.