diff options
Diffstat (limited to 'tests')
| -rw-r--r-- | tests/test_audio.py | 67 | ||||
| -rw-r--r-- | tests/test_chunking.py | 70 | ||||
| -rw-r--r-- | tests/test_tts.py | 155 |
3 files changed, 280 insertions, 12 deletions
diff --git a/tests/test_audio.py b/tests/test_audio.py index fb448f8..c043766 100644 --- a/tests/test_audio.py +++ b/tests/test_audio.py @@ -1,9 +1,11 @@ """Tests for audio helpers: speed parameters, chunk cleanup, encoding, -command construction, and duration verification.""" +command construction, duration verification, and audio concatenation.""" import tempfile import unittest +import wave from pathlib import Path +from unittest.mock import patch from converter import audio from converter import config @@ -17,6 +19,7 @@ from converter.audio import ( build_ffmetadata, build_m4b_chapters_command, cleanup_chunks, + concat_audio_files, speed_export_params, verify_output_duration, ) @@ -422,5 +425,67 @@ class BuildM4bChaptersCommandMetaTests(unittest.TestCase): self.assertNotIn("-metadata", cmd) +class ConcatAudioFilesTests(unittest.TestCase): + """Concatenation of sub-request audio into one chunk file.""" + + @staticmethod + def _write_wav(path: Path, frames: bytes, framerate: int = 24000) -> Path: + with wave.open(str(path), "wb") as wav_file: + wav_file.setnchannels(1) + wav_file.setsampwidth(2) + wav_file.setframerate(framerate) + wav_file.writeframes(frames) + return path + + def test_wav_files_are_merged_in_order(self): + with tempfile.TemporaryDirectory() as tmp: + first = self._write_wav(Path(tmp) / "a.wav", b"\x01\x00" * 10) + second = self._write_wav(Path(tmp) / "b.wav", b"\x02\x00" * 20) + destination = Path(tmp) / "out.wav" + concat_audio_files([first, second], destination) + with wave.open(str(destination), "rb") as wav_file: + self.assertEqual(wav_file.getframerate(), 24000) + self.assertEqual(wav_file.getnchannels(), 1) + self.assertEqual(wav_file.getsampwidth(), 2) + frames = wav_file.readframes(wav_file.getnframes()) + self.assertEqual(frames, b"\x01\x00" * 10 + b"\x02\x00" * 20) + + def test_single_wav_file_is_copied(self): + with tempfile.TemporaryDirectory() as tmp: + source = self._write_wav(Path(tmp) / "a.wav", b"\x03\x00" * 15) + destination = Path(tmp) / "out.wav" + concat_audio_files([source], destination) + with wave.open(str(destination), "rb") as wav_file: + self.assertEqual(wav_file.readframes(wav_file.getnframes()), + b"\x03\x00" * 15) + + def test_empty_source_list_raises(self): + with tempfile.TemporaryDirectory() as tmp: + with self.assertRaises(ValueError): + concat_audio_files([], Path(tmp) / "out.wav") + + def test_mismatched_wav_parameters_fall_back_to_ffmpeg(self): + with tempfile.TemporaryDirectory() as tmp: + first = self._write_wav(Path(tmp) / "a.wav", b"\x01\x00" * 10, framerate=24000) + second = self._write_wav(Path(tmp) / "b.wav", b"\x02\x00" * 10, framerate=16000) + destination = Path(tmp) / "out.wav" + with patch("converter.audio.shutil.which", return_value=None), \ + self.assertRaises(RuntimeError) as ctx: + concat_audio_files([first, second], destination) + self.assertIn("ffmpeg", str(ctx.exception)) + # The wave-module path must not have written a partial output. + self.assertFalse(destination.exists()) + + def test_non_wav_input_falls_back_to_ffmpeg(self): + with tempfile.TemporaryDirectory() as tmp: + source = Path(tmp) / "part.mp3" + source.write_bytes(b"not a wav file") + destination = Path(tmp) / "out.wav" + with patch("converter.audio.shutil.which", return_value=None), \ + self.assertRaises(RuntimeError) as ctx: + concat_audio_files([source], destination) + self.assertIn("ffmpeg", str(ctx.exception)) + + if __name__ == "__main__": unittest.main() diff --git a/tests/test_chunking.py b/tests/test_chunking.py index d1d95bd..2062b4e 100644 --- a/tests/test_chunking.py +++ b/tests/test_chunking.py @@ -1,18 +1,45 @@ """Tests for text chunking.""" import unittest +from unittest.mock import patch from converter import config from converter.chunking import split_into_chunks class ChunkSizeDefaultTests(unittest.TestCase): - """Guard the default chunk size: each API call is one model generation, - and long single generations lose prosody, can turn garbled, and are - truncated at the model's token limit (text past it is never spoken).""" + """Guard the request-size settings: each API call is one model + generation, and the servers silently truncate audio past their caps + (~2.5 min faster backend, ~11 min Gradio demo), so both the default + chunk size and the hard ceiling must stay well inside that budget.""" - def test_default_chunk_size_within_single_generation_budget(self): - self.assertLessEqual(config.CHUNK_SIZE_WORDS, 60) + def test_default_chunk_size_within_request_ceiling(self): + self.assertLessEqual(config.CHUNK_SIZE_WORDS, config.MAX_REQUEST_WORDS) + + def test_request_ceiling_within_single_generation_budget(self): + self.assertLessEqual(config.MAX_REQUEST_WORDS, 300) + + def test_sizes_are_positive(self): + self.assertGreaterEqual(config.CHUNK_SIZE_WORDS, 1) + self.assertGreaterEqual(config.MAX_REQUEST_WORDS, 1) + + +class RequestCeilingClampTests(unittest.TestCase): + def test_oversized_chunk_size_is_clamped_with_warning(self): + text = " ".join(f"word{i}" for i in range(30)) + "." + with patch.object(config, "MAX_REQUEST_WORDS", 10), \ + self.assertLogs("converter.chunking", level="WARNING") as logs: + chunks = split_into_chunks(text, max_words=5000) + self.assertTrue(all(len(chunk.split()) <= 10 for chunk in chunks)) + self.assertIn("clamped", " ".join(logs.output)) + + def test_default_ceiling_clamps_realistic_configuration(self): + sentences = " ".join( + f"S{i} " + " ".join(["word"] * 8) + "." for i in range(60)) + chunks = split_into_chunks(sentences, max_words=5000) + self.assertGreater(len(chunks), 1) + self.assertTrue(all(len(chunk.split()) <= config.MAX_REQUEST_WORDS + for chunk in chunks)) class SplitIntoChunksTests(unittest.TestCase): @@ -54,16 +81,39 @@ class SplitIntoChunksTests(unittest.TestCase): self.assertNotIn("12: 30", joined) def test_clause_split_requires_whitespace_after_punctuation(self): - # Run-on clauses without spaces after commas have no split point and - # must stay byte-identical rather than being re-joined with spaces. + # Run-on clauses without spaces after commas have no clause split + # point, so the last-resort word-boundary split fires instead. + # Tokens themselves (and numbers like "1,000,000") stay intact. sentence = ",".join([" ".join(["w"] * 5) for _ in range(10)]) + "." chunks = split_into_chunks(sentence, max_words=12) - self.assertEqual(chunks, [sentence]) + self.assertGreater(len(chunks), 1) + self.assertTrue(all(len(chunk.split()) <= 12 for chunk in chunks)) + tokens = sentence.replace(",", " , ").split() + rejoined = " ".join(chunks).replace(",", " , ").split() + self.assertEqual(rejoined, tokens) - def test_single_oversized_sentence_stays_intact(self): + def test_single_oversized_sentence_is_word_split(self): + # A punctuation-free sentence longer than the limit is split at word + # boundaries: the request-size ceiling is a hard limit because the + # TTS servers silently truncate oversized generations. sentence = " ".join(["word"] * 30) + "." chunks = split_into_chunks(sentence, max_words=10) - self.assertEqual(chunks, [sentence]) + self.assertGreater(len(chunks), 1) + self.assertTrue(all(len(chunk.split()) <= 10 for chunk in chunks)) + self.assertEqual(sum(len(chunk.split()) for chunk in chunks), 30) + + def test_word_split_never_breaks_number_tokens(self): + # Numbers and other punctuation-bearing tokens are single words and + # must never be broken apart by the last-resort word split. + sentence = ("There were exactly 1,000,000 soldiers marching at 12:30 " + "and " + "they kept marching onward " * 20) + "endlessly." + chunks = split_into_chunks(sentence, max_words=10) + self.assertGreater(len(chunks), 1) + joined = " ".join(chunks) + self.assertIn("1,000,000", joined) + self.assertIn("12:30", joined) + self.assertNotIn("1, 000", joined) + self.assertNotIn("12: 30", joined) if __name__ == "__main__": diff --git a/tests/test_tts.py b/tests/test_tts.py index 47a8309..26f663e 100644 --- a/tests/test_tts.py +++ b/tests/test_tts.py @@ -241,13 +241,25 @@ class FasterTTSClientGenerateTests(unittest.TestCase): sentences = [" ".join(f"word{i}" for i in range(6)) + "." for _ in range(3)] text = " ".join(sentences) pcm_parts = [b"\x01\x00" * 10, b"\x02\x00" * 20, b"\x03\x00" * 30] - with patch.object(config, "CHUNK_SIZE_WORDS", 10), \ + with patch.object(config, "MAX_REQUEST_WORDS", 10), \ patch.object(client, "_request_pcm", side_effect=pcm_parts) as mock_pcm: result = client.generate_chunk(text, 1) self.assertEqual(mock_pcm.call_count, 3) _, _, _, frames = self._read_wav(Path(result)) self.assertEqual(frames, b"".join(pcm_parts)) + def test_subchunk_size_is_clamped_to_request_ceiling(self): + client = self._make_client() + text = " ".join(f"word{i}" for i in range(8)) + pcm = b"\x01\x00" * 10 + with patch.object(config, "CHUNK_SIZE_WORDS", 4), \ + patch.object(client, "_request_pcm", return_value=pcm) as mock_pcm: + result = client.generate_chunk(text, 1) + # CHUNK_SIZE_WORDS no longer drives request size: the hard ceiling + # does, so the whole (8-word) text is one request here. + self.assertEqual(mock_pcm.call_count, 1) + self.assertIsNotNone(result) + def test_stale_chunk_files_are_removed(self): stale = Path(self._tmp.name) / "chunk_0001.mp3" stale.write_bytes(b"old") @@ -316,6 +328,147 @@ class FasterTTSClientGenerateTests(unittest.TestCase): self.assertEqual(payload["response_format"], "pcm") +class TruncationDetectionTests(unittest.TestCase): + """check_for_truncation: servers cut audio silently past their caps, so + grossly short audio must fail the request (then retry / fail visibly).""" + + def test_duration_below_ratio_raises(self): + with self.assertRaises(RuntimeError) as ctx: + tts.check_for_truncation(" ".join(["w"] * 150), 25.0, "Chunk 1") + self.assertIn("truncated", str(ctx.exception)) + + def test_duration_at_ratio_passes(self): + tts.check_for_truncation(" ".join(["w"] * 150), 30.0, "Chunk 1") + + def test_unknown_duration_skips_check(self): + tts.check_for_truncation(" ".join(["w"] * 150), None, "Chunk 1") + + def test_short_requests_skip_check(self): + tts.check_for_truncation(" ".join(["w"] * 9), 0.0, "Chunk 1") + + def test_zero_duration_fails_checked_requests(self): + with self.assertRaises(RuntimeError): + tts.check_for_truncation(" ".join(["w"] * 150), 0.0, "Chunk 1") + + +class FasterTTSClientTruncationTests(unittest.TestCase): + """Audio far shorter than its text implies fails the faster request.""" + + def setUp(self): + self._tmp = tempfile.TemporaryDirectory() + self._chunks = patch.object(tts, "CHUNKS_FOLDER", Path(self._tmp.name)) + self._chunks.start() + + def tearDown(self): + self._chunks.stop() + self._tmp.cleanup() + + def _make_client(self): + client = FasterTTSClient.__new__(FasterTTSClient) + client.voice = "default" + client.api_url = "http://127.0.0.1:8000" + return client + + def test_truncated_pcm_fails_the_chunk(self): + client = self._make_client() + text = " ".join(f"word{i}" for i in range(12)) + with patch.object(client, "_request_pcm", return_value=b"\x01\x00" * 24), \ + self.assertLogs("converter.tts", level="ERROR") as logs: + result = client.generate_chunk(text, 1) + self.assertIsNone(result) + self.assertTrue(any("truncated" in line for line in logs.output)) + + def test_full_length_pcm_passes(self): + client = self._make_client() + text = " ".join(f"word{i}" for i in range(12)) + # 12 words -> expected 4.8s, half is 2.4s -> 2.5s of audio passes. + pcm = b"\x01\x00" * int(2.5 * tts.SAMPLE_RATE) + with patch.object(client, "_request_pcm", return_value=pcm): + result = client.generate_chunk(text, 1) + self.assertIsNotNone(result) + + +class QwenTTSClientGenerateTests(unittest.TestCase): + """Qwen chunk generation: sub-request splitting and concatenation.""" + + def setUp(self): + self._tmp = tempfile.TemporaryDirectory() + self._chunks = patch.object(tts, "CHUNKS_FOLDER", Path(self._tmp.name)) + self._chunks.start() + + def tearDown(self): + self._chunks.stop() + self._tmp.cleanup() + + def _make_client(self): + client = QwenTTSClient.__new__(QwenTTSClient) + client.voice_mode = tts.VOICE_MODE_CUSTOM + return client + + @staticmethod + def _write_wav(path: Path, frames: bytes) -> Path: + with wave.open(str(path), "wb") as wav_file: + wav_file.setnchannels(1) + wav_file.setsampwidth(2) + wav_file.setframerate(tts.SAMPLE_RATE) + wav_file.writeframes(frames) + return path + + def _read_wav_frames(self, path: Path) -> bytes: + with wave.open(str(path), "rb") as wav_file: + return wav_file.readframes(wav_file.getnframes()) + + def test_single_request_copies_audio(self): + client = self._make_client() + source = self._write_wav(Path(self._tmp.name) / "server.wav", b"\x01\x00" * 50) + with patch.object(client, "_generate_custom_voice", + return_value=(str(source),)) as mock_generate: + result = client.generate_chunk("Hello world.", 1) + mock_generate.assert_called_once_with("Hello world.") + path = Path(result) + self.assertEqual(path.name, "chunk_0001.wav") + self.assertEqual(self._read_wav_frames(path), b"\x01\x00" * 50) + + def test_oversized_input_is_split_and_concatenated_in_order(self): + client = self._make_client() + first = self._write_wav(Path(self._tmp.name) / "one.wav", b"\x01\x00" * 10) + second = self._write_wav(Path(self._tmp.name) / "two.wav", b"\x02\x00" * 20) + text = " ".join(f"word{i}" for i in range(12)) + with patch.object(config, "MAX_REQUEST_WORDS", 5), \ + patch.object(client, "_generate_custom_voice", + side_effect=[(str(first),), (str(second),), + (str(first),)]) as mock_generate: + result = client.generate_chunk(text, 1) + self.assertEqual(mock_generate.call_count, 3) + path = Path(result) + self.assertEqual(path.name, "chunk_0001.wav") + self.assertEqual(self._read_wav_frames(path), + b"\x01\x00" * 10 + b"\x02\x00" * 20 + b"\x01\x00" * 10) + for call in mock_generate.call_args_list: + self.assertLessEqual(len(call[0][0].split()), 5) + + def test_truncated_sub_request_fails_the_chunk(self): + client = self._make_client() + # 0.1s of audio for 12 words (expected >= 2.4s). + source = self._write_wav(Path(self._tmp.name) / "short.wav", + b"\x01\x00" * int(0.1 * tts.SAMPLE_RATE)) + text = " ".join(f"word{i}" for i in range(12)) + with patch.object(client, "_generate_custom_voice", + return_value=(str(source),)) as mock_generate, \ + self.assertLogs("converter.tts", level="ERROR") as logs: + result = client.generate_chunk(text, 1) + self.assertIsNone(result) + self.assertEqual(mock_generate.call_count, 1) + self.assertTrue(any("truncated" in line for line in logs.output)) + + def test_empty_text_fails_the_chunk(self): + client = self._make_client() + with patch.object(client, "_generate_custom_voice") as mock_generate: + result = client.generate_chunk(" ", 1) + self.assertIsNone(result) + mock_generate.assert_not_called() + + class FasterModeWiringTests(unittest.TestCase): """AudiobookConverter wiring for the --faster backend.""" |