Use persistent HTTP clients for TTS (#125)

* Use persistent HTTP clients for TTS

Add httpx as a dependency and configure persistent HTTP clients for text-to-speech requests. The OpenAI client is now initialized with a custom httpx.Client (HTTP/2, connection limits and timeouts) in the live-stream output. The non-live output stream now reuses a requests.Session for POST requests and closes it on shutdown to improve connection reuse and performance. Also import httpx where needed and adjust resource cleanup accordingly.

* Use ANY for http_client and Session.post in tests

Update tests to accommodate an injected HTTP client and to patch the correct requests API. Added ANY to unittest.mock imports and assert_called_once_with checks for http_client=ANY in om1_speech audio live stream tests, and changed the patch target from requests.post to requests.Session.post in the audio output stream tests. These are test-only changes to align mocks with the implementation.

Author: openminddev

poetry.lock

@@ -35,6 +35,7 @@ openai= "1.60.1"
 filelock = "3.20.3"
 prometheus-client = "0.25.0"
 urllib3="2.7.0"
+httpx = {extras = ["http2"], version = "^0.28.1"}
 
 
 [tool.poetry.extras]

pyproject.toml

@@ -8,6 +8,7 @@
 from queue import Queue
 from typing import Any, Callable, Dict, Optional
 
+import httpx
 import openai
 import zenoh
 
@@ -70,6 +71,15 @@ def __init__(
         self.openai_client = openai.OpenAI(
             base_url=self._url,
             api_key=self._api_key or "no-need-api-key",
+            http_client=httpx.Client(
+                http2=True,
+                limits=httpx.Limits(
+                    max_keepalive_connections=10,
+                    max_connections=20,
+                    keepalive_expiry=300.0,
+                ),
+                timeout=httpx.Timeout(60.0, connect=5.0),
+            ),
         )
 
         # Callback for TTS state

src/om1_speech/audio/audio_output_live_stream.py

@@ -58,6 +58,9 @@ def __init__(
         # Callback for TTS state
         self._tts_state_callback = tts_state_callback
 
+        # HTTP session
+        self._session = requests.Session()
+
         # Zenoh
         self.topic = "robot/status/audio"
         self.session = None
@@ -166,7 +169,7 @@ def _process_audio(self):
             try:
                 start_time = time.time()
                 tts_request = self._pending_requests.get()
-                response = requests.post(
+                response = self._session.post(
                     self._url,
                     data=json.dumps(tts_request),
                     headers=self._headers,
@@ -381,6 +384,9 @@ def stop(self):
         """
         self.running = False
 
+        if self._session:
+            self._session.close()
+
         if self.session:
             self.session.close()
 

src/om1_speech/audio/audio_output_stream.py

@@ -3,7 +3,7 @@
 import threading
 import time
 from queue import Queue
-from unittest.mock import Mock, patch
+from unittest.mock import ANY, Mock, patch
 
 import pytest
 
@@ -518,6 +518,7 @@ def test_openai_client_creation(mock_openai, mock_zenoh):
     mock_openai.assert_called_once_with(
         base_url="http://test-server/v1",
         api_key="test-key",
+        http_client=ANY,
     )
 
     stream.stop()
@@ -534,6 +535,7 @@ def test_openai_client_no_api_key(mock_openai, mock_zenoh):
     mock_openai.assert_called_once_with(
         base_url="http://test-server/v1",
         api_key="no-need-api-key",
+        http_client=ANY,
     )
 
     stream.stop()

tests/om1_speech/audio/test_audio_output_live_stream.py

@@ -48,7 +48,7 @@ def mock_ffplay():
 
 @pytest.fixture
 def mock_requests():
-    with patch("requests.post") as mock:
+    with patch("requests.Session.post") as mock:
         # Setup mock response
         mock_response = Mock()
         mock_response.status_code = 200