Add Prometheus TTS latency metrics (#123)

* Add Prometheus TTS latency metrics

Introduce Prometheus metrics for TTS latency and record them in the TTS code paths. Adds a new src/prometheus/__init__.py exporting a Histogram (om1_tts_latency) and Gauge (om1_tts_latency_last), instruments both audio_output_stream.py and audio_output_live_stream.py to measure request duration and update the metrics, and adds prometheus-client to pyproject.toml (lockfile updated). This provides visibility into TTS processing latency for monitoring.

* Update poetry.lock content-hash

Regenerated Poetry lockfile, updating the content-hash. No other dependency or source changes were made; this reflects a lockfile refresh (e.g., after running `poetry lock`).

* Bump urllib3 to 2.7.0

Add urllib3 2.7.0 to pyproject.toml and update poetry.lock to reflect the upgrade. The lockfile bumps urllib3 from 2.6.3 to 2.7.0, updates its python-versions requirement to >=3.10, and refreshes package file hashes and the lockfile content-hash.

Author: openminddev

poetry.lock

@@ -8,7 +8,8 @@ packages = [
     { include = "om1_speech", from = "src" },
     { include = "om1_utils", from = "src" },
     { include = "om1_vlm", from = "src" },
-    { include = "zenoh_msgs", from = "src" }
+    { include = "zenoh_msgs", from = "src" },
+    { include = "prometheus", from = "src" }
 ]
 
 [tool.poetry.scripts]
@@ -33,6 +34,8 @@ av = "^15.1.0"
 vulture = "^2.14"
 openai= "1.60.1"
 filelock = "3.20.3"
+prometheus-client = "0.25.0"
+urllib3="2.7.0"
 
 
 [tool.poetry.extras]

pyproject.toml

@@ -11,6 +11,7 @@
 import openai
 import zenoh
 
+from prometheus import om1_tts_latency, om1_tts_latency_last
 from zenoh_msgs import AudioStatus, String, open_zenoh_session, prepare_header
 
 root_package_name = __name__.split(".")[0] if "." in __name__ else __name__
@@ -205,13 +206,22 @@ def _process_audio(self):
                     AudioStatus.STATUS_SPEAKER.ACTIVE.value, request_id
                 )
 
+                start_time = time.time()
                 with self.openai_client.audio.speech.with_streaming_response.create(
                     model=self._tts_model,
                     voice=self._tts_voice if tts_request["voice_id"] is None else tts_request["voice_id"],  # type: ignore
                     response_format=self._response_format,  # type: ignore
                     input=tts_request["text"],  # type: ignore
                     extra_body=self._extra_body,
                 ) as response:
+
+                    om1_tts_latency.labels(model="default").observe(
+                        time.time() - start_time
+                    )
+                    om1_tts_latency_last.labels(model="default").set(
+                        time.time() - start_time
+                    )
+
                     for chunk in response.iter_bytes(chunk_size=1024):
                         if not self.running:
                             break

src/om1_speech/audio/audio_output_live_stream.py

@@ -12,6 +12,7 @@
 import requests
 import zenoh
 
+from prometheus import om1_tts_latency, om1_tts_latency_last
 from zenoh_msgs import AudioStatus, String, open_zenoh_session, prepare_header
 
 root_package_name = __name__.split(".")[0] if "." in __name__ else __name__
@@ -163,6 +164,7 @@ def _process_audio(self):
         """
         while self.running:
             try:
+                start_time = time.time()
                 tts_request = self._pending_requests.get()
                 response = requests.post(
                     self._url,
@@ -171,6 +173,14 @@ def _process_audio(self):
                     timeout=(5, 15),
                 )
                 logger.info(f"Received TTS response: {response.status_code}")
+
+                om1_tts_latency.labels(model="default").observe(
+                    time.time() - start_time
+                )
+                om1_tts_latency_last.labels(model="default").set(
+                    time.time() - start_time
+                )
+
                 if response.status_code == 200:
                     result = response.json()
                     if "response" in result:

src/om1_speech/audio/audio_output_stream.py

@@ -0,0 +1,11 @@
+from prometheus_client import Gauge, Histogram
+
+om1_tts_latency = Histogram(
+    "om1_tts_latency_seconds", "Latency of TTS processing in seconds", ["model"]
+)
+
+om1_tts_latency_last = Gauge(
+    "om1_tts_latency_last_seconds",
+    "Latency of the last TTS processing in seconds",
+    ["model"],
+)