"metrics": { "total_requests": 3, "total_duration": 15.774682521820068, "total_ttft": 0.8553550243377686, "avg_ttft": 0.28511834144592285, "total_input_tokens": 12144, "total_output_tokens": 66, "total_tokens": 12210, "avg_tokens_per_second": 4.183919385300249, "total_audio_input_tokens": 3, "total_text_input_tokens": 12141, "total_image_input_tokens": 0, "total_cached_tokens": 0, "total_text_output_tokens": 66, "total_audio_output_tokens": 0, "total_image_output_tokens": 0, "cancelled_requests": 0 } above are the sum metrices of the session now i am using gemini live api how to calculate the cost the rates are different for text input text output and audio input and audio output how will calculate the cost here i dont see the audio output here

Thanks for reaching out, Aman. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

i dont see the audio output tokens .. even though i had a audio session with the model via livekit voice agent

The absence of audio output tokens in your metrics likely means that the Gemini Live API didn’t include audio-specific token details in its response. The LiveKit metrics system is set up to capture audio output tokens, but it relies on the API providing that information. Here’s a gentle way to check this further: 1. Take a look at the

usage_metadata

from the API response and see if there’s an entry in

response_tokens_details

where the

modality

is marked as

AUDIO

. 2. If you don’t find such an entry, it probably means the API isn’t categorizing or reporting audio output tokens. If you find that the API response is missing this data, it might be worth reaching out for further investigation or clarification from the API provider. Please feel free to ask if you'd like a hand inspecting the

usage_metadata

or if you have any other questions—I’m here to help! Sources: livekit-agents/livekit/agents/metrics/utils.py | plugins/openai/src/realtime/api_proto.ts

currently i am using @session.on("metrics_collected") def _on_metrics_collected(ev: MetricsCollectedEvent): session_metrics.add_metric(ev.metrics) to collect the metrics give me code to get the usage_metadata

The

@session.on("metrics_collected")

event doesn’t give direct access to the raw

usage_metadata

. Instead, it provides structured metrics like

RealtimeModelMetrics

or

LLMMetrics

, which are thoughtfully derived from the

usage_metadata

. Here’s a friendly example of how you can work with the available metrics:

python @session.on("metrics_collected") def _on_metrics_collected(ev: MetricsCollectedEvent): m = ev.metrics if isinstance(m, RealtimeModelMetrics): print(f"Input Tokens: {m.input_tokens}, Output Tokens: {m.output_tokens}") print(f"Audio Input Tokens: {m.input_token_details.audio_tokens}") print(f"Text Input Tokens: {m.input_token_details.text_tokens}") print(f"Audio Output Tokens: {m.output_token_details.audio_tokens}") print(f"Text Output Tokens: {m.output_token_details.text_tokens}")

If you’re looking for the raw

usage_metadata

, it’s not directly available through the public API. But no worries! You can enable OpenTelemetry tracing to capture detailed trace data, including token usage and other helpful attributes. I’m here to help if you want assistance setting up tracing or if you have any other questions — just let me know! Sources: examples/voice_agents/langfuse_trace.py | plugins/google/src/beta/realtime/realtime_api.ts | Logs, metrics, and telemetry | LiveKit Docs