https://livekit.io logo
Join Slack
Powered by
<#C088ZNU7QQ5> how to excatly extract the metircs ...
# ask-ai
s
#C088ZNU7QQ5 how to excatly extract the metircs of stt where i am using streaming api of sarvam ai model
t
Thanks for reaching out, Karthik. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.
Thanks — I’m responding to your question about how to exactly extract STT metrics when using the Sarvam AI streaming API. Short answer: use the Agents framework’s event-based metrics system. Attach listeners to the STT component to receive 
metrics_collected
and 
eou_metrics_collected
events; those events contain the STT and end-of-utterance metrics you need. If you want a runnable example, please tell me whether you’re using Python or Node and whether you want metrics logged server-side (inside an AgentSession) or forwarded to your own telemetry system. Below is a concise Python example and a quick explanation of the key metrics. 1) Minimal Python example (inside your AgentSession)
Copy code
from livekit.plugins import sarvam
from livekit.agents import AgentSession

# create session with Sarvam STT
session = AgentSession(
    stt=sarvam.STT(model="saaras:v3", language="hi-IN"),
    # ... other configuration
)

# STT metrics listener
def on_stt_metrics(metrics):
    print(f"STT duration: {metrics.duration:.4f}s")
    print(f"Audio duration: {metrics.audio_duration:.4f}s")
    print(f"Speech ID: {metrics.speech_id}")
    print(f"Error: {metrics.error}")
    print(f"Streamed: {metrics.streamed}")

# End-of-utterance metrics listener
def on_eou_metrics(metrics):
    print(f"End-of-utterance delay: {metrics.end_of_utterance_delay:.4f}s")
    print(f"Transcription delay: {metrics.transcription_delay:.4f}s")
    print(f"Speech ID: {metrics.speech_id}")

# Attach listeners (do this during session startup / on_enter)
session.stt.on("metrics_collected", on_stt_metrics)
session.stt.on("eou_metrics_collected", on_eou_metrics)
2) Key metrics you’ll receive • duration: total processing time for the STT request. • audio_duration: length of the audio input in seconds. • end_of_utterance_delay: time from speech end to EOU detection. • transcription_delay: time from EOU to full transcription. • speech_id: unique identifier for the speech segment. • error: any error message when transcription fails. • streamed: boolean indicating whether the response was streamed. 3) Where to place listeners • Attach these listeners in your agent’s startup flow (for example, inside 
on_enter()
), so they activate when the session begins. Next steps / quick questions to help me tailor this: • Do you want sample code in Node instead of Python? • Do you need the metrics sent to a logging/monitoring system (e.g., Prometheus, Datadog) or simply printed/logged server-side? • Which specific metric(s) are most important for your use case (latency, EOU timings, audio length, errors, etc.)? Tell me your answers and I’ll give the exact sample code/config for your environment. Sources: Sarvam STT plugin guide | LiveKit Documentation | STT Metrics | LiveKit Documentation | Speech-to-text (STT) models overview | LiveKit Documentation
s
Right now we're listening on session.on("metrics_collected") which gives us the generic MetricsCollectedEvent, i am getting in the metrics as 0.00s and all. so i want accurate numbers. what can i do?
t
Thanks — I hear you: you’re receiving generic zero-valued metrics from 
session.on("metrics_collected")
and need accurate STT timings. Let’s narrow this down and get you precise numbers. Quick checklist (please confirm these so I can give exact code): • Are you using Python or Node? • Which LiveKit Agents / livekit-plugins-sarvam version are you running? • Are you attaching the listener on the top-level session (
session.on("metrics_collected")
) or on the STT component (
session.stt.on(...)
)? • Are you running the listeners inside the agent’s startup (for example, inside 
on_enter()
), or after the session has already started? Why you’re likely seeing zeros • The generic 
session.on("metrics_collected")
receives a broad MetricsCollectedEvent that may not include the detailed STT-specific timing fields (or they may be unpopulated) — this can yield zeros. • STT providers emit richer, provider-specific events (like 
metrics_collected
and 
eou_metrics_collected
) on the STT component itself. Those events include 
duration
, 
audio_duration
, 
end_of_utterance_delay
, 
transcription_delay
, 
speech_id
, 
streamed
, and 
error
. Listening on the STT component is required to receive accurate values. What to do to get accurate metrics 1. Attach listeners directly to the STT plugin instead of the top-level session event: ◦ Use 
session.stt.on("metrics_collected", handler)
and 
session.stt.on("eou_metrics_collected", handler)
so you receive STT-specific metric objects. 1. Ensure listeners are registered before streaming begins: ◦ Add the listeners during agent startup (for example, 
on_enter()
), so they’re active for all incoming audio. 1. Confirm the Sarvam STT plugin is configured and enabled: ◦ Make sure you initialized 
sarvam.STT(...)
in your AgentSession and that 
SARVAM_API_KEY
is set. 1. Verify you are reading the right fields from the STT metrics object: ◦ Use 
metrics.duration
, 
metrics.audio_duration
, 
metrics.speech_id
, and for EOU events 
metrics.end_of_utterance_delay
and 
metrics.transcription_delay
. 1. Check for streamed vs non-streamed behavior: ◦ If the STT response is streamed, some timing values may be populated differently. The 
streamed
boolean on the STT metrics helps interpret them. 1. If values are still zero: ◦ Confirm audio is actually being delivered to the STT plugin (log incoming audio timestamps). ◦ Confirm the STT provider returns metrics for your chosen model (
saaras:v3
etc.). Some models/providers may not populate all fields. ◦ Ensure there are no exceptions in the STT pipeline that cause early returns — inspect 
metrics.error
. Minimal Python example (attach to STT component)
Copy code
# Attach listeners during startup / on_enter
def on_stt_metrics(metrics):
    print("STT metrics:", metrics.duration, metrics.audio_duration, metrics.streamed, metrics.error)

def on_eou_metrics(metrics):
    print("EOU metrics:", metrics.end_of_utterance_delay, metrics.transcription_delay, metrics.speech_id)

session.stt.on("metrics_collected", on_stt_metrics)
session.stt.on("eou_metrics_collected", on_eou_metrics)
Next steps I can provide once you confirm environment: • A Node or Python full example placed in the correct lifecycle hook (e.g., 
on_enter()
). • Targeted debug steps to confirm audio delivery and verify Sarvam is populating metrics. • A snippet that forwards metrics to your logging/monitoring so you can audit values per request. Tell me: which language and where you’re currently attaching the listener (session vs session.stt), plus your plugin version — I’ll give the exact code and debugging steps.
2 Views
Open in Slack
PreviousNext
Powered by