ops: Add OTEL metric for STT latency

app/helpers/call_llm.py

@@ -39,6 +39,7 @@
     SpanAttributeEnum,
     call_answer_latency,
     call_cutoff_latency,
+    call_stt_complete_latency,
     gauge_set,
     tracer,
 )
@@ -233,6 +234,17 @@ async def _commit_answer(
             if wait:
                 await last_chat
 
+        async def _compute_stt_metrics() -> None:
+            """
+            Report the recognition latency.
+            """
+            start = time.monotonic()
+            await stt_complete_gate.wait()
+            gauge_set(
+                metric=call_stt_complete_latency,
+                value=time.monotonic() - start,
+            )
+
         async def _response_callback(_retry: bool = False) -> None:
             """
             Triggered when the audio buffer needs to be processed.
@@ -243,6 +255,9 @@ async def _response_callback(_retry: bool = False) -> None:
             nonlocal answer_start
             answer_start = time.monotonic()
 
+            # Report the STT metrics
+            stt_metrics_task = asyncio.create_task(_compute_stt_metrics())
+
             # Wait the complete recognition for 50ms maximum
             try:
                 await asyncio.wait_for(stt_complete_gate.wait(), timeout=0.05)
@@ -277,8 +292,11 @@ async def _response_callback(_retry: bool = False) -> None:
                     )
                 )
 
-            # Process the response
-            await _commit_answer(wait=True)
+            # Process the response and wait for latency metrics
+            await asyncio.gather(
+                _commit_answer(wait=False),
+                stt_metrics_task,
+            )
 
         # First call
         if len(call.messages) <= 1:

app/helpers/monitoring.py

@@ -67,6 +67,8 @@ class SpanMeterEnum(str, Enum):
     """Audio frames in latency in seconds."""
     CALL_FRAMES_OUT_LATENCY = "call.frames.out.latency"
     """Audio frames out latency in seconds."""
+    CALL_STT_COMPLETE_LATENCY = "call.stt.complete.latency"
+    """Speech-to-text missed complete latency."""
 
     def counter(
         self,
@@ -127,6 +129,7 @@ def gauge(
 call_cutoff_latency = SpanMeterEnum.CALL_CUTOFF_LATENCY.gauge("s")
 call_frames_in_latency = SpanMeterEnum.CALL_FRAMES_IN_LATENCY.gauge("s")
 call_frames_out_latency = SpanMeterEnum.CALL_FRAMES_OUT_LATENCY.gauge("s")
+call_stt_complete_latency = SpanMeterEnum.CALL_STT_COMPLETE_LATENCY.gauge("s")
 
 
 def gauge_set(