Feat/voice robustness

examples/voice_turn_modes.py

@@ -18,21 +18,41 @@
 
 from __future__ import annotations
 
+import asyncio
 import os
+import random
+from datetime import datetime, timezone
 
 from timbal import Agent
+from timbal.core.tool import Tool
 from timbal.voice import resolve_turn_detector
 
 _MODE = os.environ.get("TIMBAL_VOICE_TURN_DETECTOR", "heuristic").strip().lower()
 
+
+async def get_datetime() -> str:
+    """Get the current UTC date and time (slow on purpose for voice tool-call testing)."""
+    await asyncio.sleep(random.uniform(3.0, 5.0))
+    return datetime.now(timezone.utc).strftime("%Y-%m-%d %H:%M:%S %Z")
+
+
 agent = Agent(
     name="voice_turn_modes",
-    model=os.environ.get("TIMBAL_VOICE_DEMO_MODEL", "openai/gpt-4o-mini"),
+    model=os.environ.get("TIMBAL_VOICE_DEMO_MODEL", "groq/llama-3.1-8b-instant"),
+    max_tokens=1024,  # required when switching to Anthropic from the playground
     system_prompt=(
         "You are a concise voice assistant. Keep replies to 1–2 short sentences. "
+        "When the user asks for the date and/or time, you MUST call get_datetime "
+        "and then speak the returned value clearly (include both date and time). "
+        "Never claim you cannot tell the time. "
         f"(turn_detector mode: {_MODE})"
     ),
-    tools=[],
+    tools=[
+        Tool(
+            handler=get_datetime,
+            description="Return the current UTC date and time. Always use this for date/time questions.",
+        )
+    ],
 )
 
 # Server reads this on startup (instance or mode name). Client WS JSON cannot override it.

pyproject.toml

@@ -63,6 +63,7 @@ server = [
 voice = [
   "huggingface-hub>=0.28.0",
   "onnxruntime>=1.20.0",
+  "transformers>=4.40.0",
 ]
 all = [
   "timbal[codegen]",

python/tests/server/test_voice_config.py

@@ -24,6 +24,7 @@
 class TestDefaultVoiceConfigFromEnv:
     def test_defaults_when_unset(self) -> None:
         cfg = voice_routes.default_voice_config_from_env()
+        assert cfg["stt_provider"] == "elevenlabs"
         assert cfg["stt_model"] == "scribe_v2_realtime"
         assert cfg["tts_model"] == "eleven_flash_v2_5"
         assert cfg["voice"] == voice_routes._DEFAULT_VOICE_ID
@@ -33,10 +34,12 @@ def test_defaults_when_unset(self) -> None:
         assert cfg["tts_extra"]["auto_mode"] is True
 
     def test_env_overrides(self, monkeypatch: pytest.MonkeyPatch) -> None:
+        monkeypatch.setenv("TIMBAL_STT_PROVIDER", "deepgram")
         monkeypatch.setenv("TIMBAL_STT_MODEL", "custom_stt")
         monkeypatch.setenv("TIMBAL_TTS_MODEL", "custom_tts")
         monkeypatch.setenv("TIMBAL_VOICE_LANGUAGE", "en")
         cfg = voice_routes.default_voice_config_from_env()
+        assert cfg["stt_provider"] == "deepgram"
         assert cfg["stt_model"] == "custom_stt"
         assert cfg["tts_model"] == "custom_tts"
         assert cfg["language"] == "en"

python/tests/server/voice_env.py

@@ -4,6 +4,7 @@
 
 # Env vars that affect ``default_voice_config_from_env`` / ``merge_voice_config``.
 VOICE_ENV_KEYS = (
+    "TIMBAL_STT_PROVIDER",
     "TIMBAL_STT_MODEL",
     "TIMBAL_TTS_MODEL",
     "ELEVENLABS_VOICE_ID",

python/tests/utils/test_serialization.py

@@ -169,6 +169,16 @@ def test_empty_string_returns_empty_dict(self):
         assert coerce_to_dict("") == {}
         assert coerce_to_dict("   ") == {}
 
+    def test_nullish_no_arg_tool_inputs(self):
+        # Groq / OpenAI-compat often emit null for tools with zero parameters.
+        assert coerce_to_dict(None) == {}
+        assert coerce_to_dict("null") == {}
+        assert coerce_to_dict("None") == {}
+        assert coerce_to_dict("NULL") == {}
+
+    def test_json_null_literal_returns_empty_dict(self):
+        assert coerce_to_dict("null") == {}
+
     def test_literal_eval_fallback(self):
         # Python dict literal that isn't valid JSON (single quotes)
         assert coerce_to_dict("{'a': 1}") == {"a": 1}
@@ -177,6 +187,10 @@ def test_unparseable_string_raises(self):
         with pytest.raises(ValueError, match="Cannot coerce value to dict"):
             coerce_to_dict("not a dict at all!")
 
+    def test_non_dict_json_raises(self):
+        with pytest.raises(ValueError, match="Cannot coerce value to dict"):
+            coerce_to_dict("[1, 2, 3]")
+
     def test_non_string_non_dict_raises(self):
         with pytest.raises(ValueError, match="Cannot coerce value to dict"):
             coerce_to_dict(42)