[gemini-3.5-transcribe] AudioTranscriptionConfig.diarization=True is silently ignored without returning speaker labels

Summary

In the google-genai Python SDK, types.AudioTranscriptionConfig explicitly declares the diarization: bool field ("Configures speaker diarization").

However, when passing diarization=True to client.interactions.create(model="gemini-3.5-transcribe", ...) with mode="verbatim", the request completes successfully with status completed and word-level timestamps, but speaker diarization is completely silently ignored. None of the returned word/segment annotations contain speaker_label or speaker fields.

Environment Details

OS: macOS (Darwin arm64)
Python: 3.13
SDK: google-genai (latest)
Model: gemini-3.5-transcribe

Minimal Reproducible Example

from google import genai

client = genai.Client()
audio_file = client.files.upload(file="two_speaker_interview.wav")

interaction = client.interactions.create(
    model="gemini-3.5-transcribe",
    input=[{
        "type": "audio",
        "uri": audio_file.uri,
        "mime_type": "audio/wav",
    }],
    generation_config={
        "transcription_config": {
            "mode": {"type": "verbatim"},
            "timestamp_granularities": ["word"],
            "diarization": True,  # Explicitly requested
        }
    },
)

steps = getattr(interaction, "steps", [])
speakers = set()
for step in steps:
    for content in getattr(step, "content", []):
        for ann in getattr(content, "annotations", []):
            spk = ann.get("speaker_label") or ann.get("speaker")
            if spk:
                speakers.add(spk)

print(f"Status: {interaction.status}")
print(f"Total words: {len(steps)}")
print(f"Detected speakers: {speakers}")
# Output:
# Status: completed
# Total words: 274
# Detected speakers: set()

Actual Behavior

The API succeeds with status completed, but:

Zero speaker labels (speaker_label or speaker) are populated in the annotations.
The API silently discards diarization=True without any warning or validation error.

Expected Behavior

Either populate speaker_label (spk_1, spk_2, etc.) in the annotations as defined in the SDK data schema;
Or raise a validation/not-implemented error informing the caller that diarization is not supported for gemini-3.5-transcribe.

Additional Context

Related to earlier report #2923. Unlike earlier versions where requests hung on incomplete, the interaction now completes, but the diarization functionality is dropped silently.

@randyy179

Hi @randyy179, thank you for reporting this issue and providing a clear reproduction snippet!
We have reproduced this behavior where diarization=True in transcription_config for gemini-3.5-transcribe completes with status="completed" but returns no speaker labels (Detected speakers: set()), and we have escalated it to our engineering team.
Thanks

@randyy179

HI @randyy179

Thanks for the detailed report and repro, this was really helpful.

AudioTranscriptionConfig.diarization (a bool) is the field for the generate_content API. The Interactions API (client.interactions.create) uses a different shape: diarization_mode="speaker", nested inside the transcription mode:

Docs: https://ai.google.dev/gemini-api/docs/transcribe#speaker-diarization

Why there was no error: you're right that this should have failed loudly. Right now the SDK accepts unknown fields like diarization and then drops them before the request is sent, so the API never sees the field and can't reject it. That's a bug on our side, and we're tracking a fix so unknown fields are sent to the API and you get a clear error.

In the meantime while we patch the error behavior, hope diarization_mode unblocks you

feat(python): preserve unknown fields and open-union variants in Inte…

fix: preserve unknown fields and open-union variants in Interactions …

Navigation Menu

Repository navigation

[gemini-3.5-transcribe] AudioTranscriptionConfig.diarization=True is silently ignored without returning speaker labels #3162

Summary

Environment Details

Minimal Reproducible Example

Actual Behavior

Expected Behavior

Additional Context

2 remaining items

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

[gemini-3.5-transcribe] AudioTranscriptionConfig.diarization=True is silently ignored without returning speaker labels #3162

Description

Summary

Environment Details

Minimal Reproducible Example

Actual Behavior

Expected Behavior

Additional Context

Activity

Venkaiahbabuneelam commented on Oct 19, 2026

treyn-create commented on Oct 21, 2026

2 remaining items

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions