# # Copyright (c) 2024-2026, Daily # # SPDX-License-Identifier: BSD 2-Clause License # """Tests for the eval service constructors (config -> EvalJudge/EvalSpeech/EvalTranscriber).""" import unittest from pipecat.evals.judge import EvalJudge from pipecat.evals.services import _cfg_language, cartesia_service from pipecat.evals.speech import EvalSpeech, tts_cache_key, tts_sample_rate from pipecat.evals.transcribe import EvalTranscriber from pipecat.transcriptions.language import Language from pipecat.utils.types import NOT_GIVEN def _fake_stt(config, sample_rate): return ("FAKE_STT", config, sample_rate) def _fake_tts(config, sample_rate): return ("FAKE_TTS", config, sample_rate) def _fake_judge_llm(config): return ("FAKE_JUDGE", config) class TestTranscriberFromConfig(unittest.TestCase): def test_unknown_service_rejected(self): with self.assertRaises(ValueError): EvalTranscriber.from_config({"service": "nope"}) def test_factory_escape_hatch(self): t = EvalTranscriber.from_config({"factory": "tests.test_evals_services._fake_stt"}) self.assertEqual(t._service[0], "FAKE_STT") self.assertEqual(t._service[2], 16000) # STT_SAMPLE_RATE def test_padding_secs(self): from pipecat.evals.transcribe import SILENCE_PAD_S default = EvalTranscriber.from_config({"factory": "tests.test_evals_services._fake_stt"}) self.assertEqual(default._padding_secs, SILENCE_PAD_S) override = EvalTranscriber.from_config( {"factory": "tests.test_evals_services._fake_stt", "padding_secs": 0.5} ) self.assertEqual(override._padding_secs, 0.5) class TestVoiceFromConfig(unittest.TestCase): def test_cache_key_excludes_sample_rate(self): a = tts_cache_key({"service": "kokoro", "voice": "v", "model": "m", "sample_rate": 16000}) b = tts_cache_key({"service": "kokoro", "voice": "v", "model": "m", "sample_rate": 24000}) self.assertEqual(a, b) def test_cache_key_distinguishes_voice(self): self.assertNotEqual( tts_cache_key({"service": "kokoro", "voice": "a"}), tts_cache_key({"service": "kokoro", "voice": "b"}), ) def test_cache_key_distinguishes_language(self): # Two configs identical except for language must not collide, so an # English and a Chinese render of the same text get separate cache slots. self.assertNotEqual( tts_cache_key({"service": "cartesia", "voice": "v", "language": "en"}), tts_cache_key({"service": "cartesia", "voice": "v", "language": "zh"}), ) # An absent language and an explicit empty one key to the same slot. self.assertEqual( tts_cache_key({"service": "cartesia", "voice": "v"}), tts_cache_key({"service": "cartesia", "voice": "v", "language": ""}), ) def test_sample_rate_default(self): self.assertEqual(tts_sample_rate({}), 16000) self.assertEqual(tts_sample_rate({"sample_rate": 24000}), 24000) def test_unknown_service_rejected(self): with self.assertRaises(ValueError): EvalSpeech.from_config({"service": "nope", "voice": "v"}) def test_missing_service_or_voice_rejected(self): with self.assertRaises(ValueError): EvalSpeech.from_config({}) def test_factory_escape_hatch(self): v = EvalSpeech.from_config( {"factory": "tests.test_evals_services._fake_tts", "sample_rate": 24000} ) self.assertEqual(v._service[0], "FAKE_TTS") self.assertEqual(v._service[2], 24000) def test_language_reaches_cartesia_settings(self): # Cartesia is the one builder a unit test can construct: Whisper, Moonshine # and Kokoro load their models at construction time. service = cartesia_service( {"service": "cartesia", "voice": "v", "api_key": "test-key", "language": "zh"}, 16000, ) self.assertEqual(service._settings.language, Language.ZH) def test_no_language_leaves_cartesia_default(self): # Omitting language must not force a value; the service keeps its own # default, which for Cartesia is Language.EN. service = cartesia_service( {"service": "cartesia", "voice": "v", "api_key": "test-key"}, 16000 ) self.assertEqual(service._settings.language, Language.EN) def test_websocket_service_rejected(self): # run_tts can't be driven without a pipeline to manage the connection, so a # websocket-streaming TTS service must be rejected at construction. from pipecat.services.websocket_service import WebsocketService class _FakeWS(WebsocketService): async def _connect_websocket(self): pass async def _disconnect_websocket(self): pass async def _receive_messages(self): pass with self.assertRaises(ValueError): EvalSpeech(_FakeWS(), sample_rate=16000, cache_key="k") class TestCfgLanguage(unittest.TestCase): def test_absent_leaves_the_field_unset(self): self.assertIs(_cfg_language({}), NOT_GIVEN) self.assertIs(_cfg_language({"language": None}), NOT_GIVEN) def test_blank_leaves_the_field_unset(self): # A key present but empty in the YAML means "unset", not "unknown language". self.assertIs(_cfg_language({"language": ""}), NOT_GIVEN) self.assertIs(_cfg_language({"language": " "}), NOT_GIVEN) def test_code_or_language_accepted(self): self.assertEqual(_cfg_language({"language": "zh"}), Language.ZH) self.assertEqual(_cfg_language({"language": " zh-TW "}), Language.ZH_TW) self.assertEqual(_cfg_language({"language": Language.ES}), Language.ES) def test_unknown_code_rejected(self): with self.assertRaises(ValueError): _cfg_language({"language": "notalang"}) def test_non_string_rejected(self): # YAML 1.1 reads a bare `language: no` as False rather than Norwegian, so # the coercion has to reject non-strings instead of passing them through. with self.assertRaises(ValueError): _cfg_language({"language": False}) class _CountingTTS: """Minimal stand-in for a TTSService: run_tts yields one audio frame.""" def __init__(self, pcm: bytes, sample_rate: int): self.pcm = pcm self.sample_rate = sample_rate self.calls = 0 async def run_tts(self, text, context_id): from pipecat.frames.frames import TTSAudioRawFrame self.calls += 1 yield TTSAudioRawFrame(audio=self.pcm, sample_rate=self.sample_rate, num_channels=1) class TestSpeechCache(unittest.IsolatedAsyncioTestCase): async def test_cache_round_trip_and_sr_mismatch(self): import tempfile pcm = b"\x01\x02" * 1600 # 100ms of 16kHz mono with tempfile.TemporaryDirectory() as tmp: tts = _CountingTTS(pcm, 16000) speech = EvalSpeech(tts, sample_rate=16000, cache_key="k", cache_dir=tmp) speech._started = True # skip the FrameProcessor lifecycle out, sr = await speech.generate("hello") self.assertEqual((out, sr), (pcm, 16000)) self.assertEqual(tts.calls, 1) # Second call hits the WAV cache; the service is not called again. out2, _ = await speech.generate("hello") self.assertEqual(out2, pcm) self.assertEqual(tts.calls, 1) # A different requested sample rate misses the cached file's rate and # regenerates (the cache slot is shared across rates by design). tts24 = _CountingTTS(pcm, 24000) speech24 = EvalSpeech(tts24, sample_rate=24000, cache_key="k", cache_dir=tmp) speech24._started = True await speech24.generate("hello") self.assertEqual(tts24.calls, 1) class TestJudgeFromConfig(unittest.TestCase): def test_unknown_service_rejected(self): with self.assertRaises(ValueError): EvalJudge.from_config({"service": "nope"}) def test_factory_escape_hatch(self): j = EvalJudge.from_config({"factory": "tests.test_evals_services._fake_judge_llm"}) self.assertIsNotNone(j) self.assertEqual(j._service[0], "FAKE_JUDGE") if __name__ == "__main__": unittest.main()