mirror of
https://github.com/yusufipk/dikte.git
synced 2026-09-11 19:06:11 +00:00
yusufipk
2cfbbb2d99
whisper-server is started on --inference-path /v1/audio/transcriptions, which is exactly the path api.py already builds for the hosted providers, and llama-server answers /chat/completions the way OpenRouter does. So the local half is one more base URL rather than a second code path: worker.py, filetranscribe.py and meeting.py are untouched, and dictation, subtitles and meetings all work here on the first try. Three findings worth naming, none of them in the new code: whisper.cpp cuts segments on tokens, which in Turkish lands inside a word about as often as between two. Pasted raw that gives "akraba değ\niller."; in a subtitle it gives a cue reading "değ". Whisper marks the start of a word with a leading space, so a piece that does not begin with one continues the word above it. A small model will repeat the transcript until the context is full, and every one of those tokens is a second of somebody waiting: measured at 206 seconds, and 25 with a ceiling on the reply. Hosted models are left alone, where the same runaway is rare and a ceiling would cut the minutes short. A server outlives SIGTERM and SIGKILL holding its model in memory. Signals are now turned into an event Qt delivers, since Qt blocks in C where a Python handler never runs, and a pid file lets the next start sweep up what a SIGKILL left behind. The minutes keep their own provider rather than following cleanup's. The two jobs are not the same size: a 4B model here will strip the filler words out of a dictation and will not write up an hour long meeting. The suite runs offline now: a test that reaches the network says so instead of quietly going there.
238 lines
9.8 KiB
Python
238 lines
9.8 KiB
Python
"""Transcribing a file: splitting it, stamping it, and writing subtitles.
|
|
|
|
to_srt is the awkward one. The text is the authority on wording and the segments
|
|
on timing, and they meet at a whole-second stamp that a cleanup model was asked
|
|
to leave alone. It has to survive a model that wrapped a line, dropped one, or
|
|
made up a stamp nobody recorded.
|
|
"""
|
|
|
|
import contextlib
|
|
import unittest
|
|
import wave
|
|
from unittest import mock
|
|
|
|
import api
|
|
import filetranscribe as ft
|
|
from tests.support import DikteTest, make_wav, silence, tone
|
|
|
|
|
|
class Timestamps(unittest.TestCase):
|
|
def test_under_an_hour(self):
|
|
self.assertEqual(ft.format_timestamp(0), "00:00")
|
|
self.assertEqual(ft.format_timestamp(65.9), "01:05")
|
|
self.assertEqual(ft.format_timestamp(599), "09:59")
|
|
|
|
def test_past_an_hour_the_hours_show(self):
|
|
self.assertEqual(ft.format_timestamp(3600), "1:00:00")
|
|
self.assertEqual(ft.format_timestamp(3725), "1:02:05")
|
|
|
|
def test_srt_wants_milliseconds_and_a_comma(self):
|
|
self.assertEqual(ft.srt_timestamp(0), "00:00:00,000")
|
|
self.assertEqual(ft.srt_timestamp(1.5), "00:00:01,500")
|
|
self.assertEqual(ft.srt_timestamp(3725.25), "01:02:05,250")
|
|
|
|
def test_a_negative_start_is_pulled_up_to_zero(self):
|
|
self.assertEqual(ft.srt_timestamp(-3), "00:00:00,000")
|
|
|
|
|
|
class ToSrt(unittest.TestCase):
|
|
def test_nothing_to_do(self):
|
|
self.assertEqual(ft.to_srt("", []), "")
|
|
self.assertEqual(ft.to_srt("no stamps here", []), "")
|
|
|
|
def test_a_cue_takes_its_timing_from_the_segment_it_came_from(self):
|
|
srt = ft.to_srt("[00:01] Hello there.", [(1.25, 2.75, "hello there")])
|
|
self.assertIn("00:00:01,250 --> 00:00:02,750", srt)
|
|
self.assertIn("Hello there.", srt)
|
|
|
|
def test_the_text_wins_on_wording(self):
|
|
"""Cleanup edits survive; only the timing comes from the segments."""
|
|
srt = ft.to_srt("[00:01] Hello there.", [(1.0, 2.0, "uh hello uh there")])
|
|
self.assertIn("Hello there.", srt)
|
|
self.assertNotIn("uh", srt)
|
|
|
|
def test_cues_are_numbered_from_one(self):
|
|
srt = ft.to_srt("[00:00] One\n[00:02] Two",
|
|
[(0.0, 1.0, "One"), (2.0, 3.0, "Two")])
|
|
self.assertTrue(srt.startswith("1\n"))
|
|
self.assertIn("\n2\n", srt)
|
|
|
|
def test_a_wrapped_line_joins_the_cue_above_it(self):
|
|
srt = ft.to_srt("[00:01] Hello there,\nand welcome.",
|
|
[(1.0, 4.0, "hello there and welcome")])
|
|
self.assertIn("Hello there, and welcome.", srt)
|
|
self.assertEqual(srt.count(" --> "), 1)
|
|
|
|
def test_a_stamp_nobody_recorded_still_gets_timing(self):
|
|
srt = ft.to_srt("[00:05] Invented.", [])
|
|
self.assertIn("00:00:05,000 --> ", srt)
|
|
|
|
def test_a_cue_with_no_end_runs_until_the_next_one(self):
|
|
srt = ft.to_srt("[00:00] One\n[00:04] Two", [])
|
|
self.assertIn("00:00:00,000 --> 00:00:04,000", srt)
|
|
|
|
def test_the_last_cue_gets_a_minimum_length(self):
|
|
srt = ft.to_srt("[00:10] Last words.", [])
|
|
self.assertIn("00:00:10,000 --> 00:00:11,500", srt)
|
|
|
|
def test_a_cue_is_cut_short_when_the_next_one_starts_first(self):
|
|
"""Whisper's end times overlap now and then; subtitles must not."""
|
|
srt = ft.to_srt("[00:00] One\n[00:02] Two",
|
|
[(0.0, 9.0, "One"), (2.0, 3.0, "Two")])
|
|
self.assertIn("00:00:00,000 --> 00:00:02,000", srt)
|
|
|
|
def test_blank_lines_and_empty_cues_are_dropped(self):
|
|
srt = ft.to_srt("[00:00] One\n\n[00:02]\n[00:03] Three", [])
|
|
self.assertEqual(srt.count(" --> "), 2)
|
|
|
|
def test_the_hour_form_of_a_stamp_is_understood(self):
|
|
srt = ft.to_srt("[1:02:05] Late.", [])
|
|
self.assertIn("01:02:05,000", srt)
|
|
|
|
def test_the_file_ends_with_a_newline(self):
|
|
self.assertTrue(ft.to_srt("[00:00] One", []).endswith("\n"))
|
|
|
|
|
|
class SplitText(unittest.TestCase):
|
|
def test_short_text_stays_whole(self):
|
|
self.assertEqual(ft.split_text("hello", False), ["hello"])
|
|
|
|
def test_a_long_transcript_is_broken_up(self):
|
|
text = " ".join(["word"] * 8000)
|
|
blocks = ft.split_text(text, False)
|
|
self.assertGreater(len(blocks), 1)
|
|
for block in blocks:
|
|
self.assertLessEqual(len(block), ft.CLEANUP_CHUNK_CHARS)
|
|
|
|
def test_nothing_is_lost_in_the_splitting(self):
|
|
text = " ".join(f"word{index}" for index in range(4000))
|
|
self.assertEqual(" ".join(ft.split_text(text, False)), text)
|
|
|
|
def test_a_timestamped_transcript_is_never_broken_mid_line(self):
|
|
text = "\n".join(f"[00:{index:02d}] a line of some length here"
|
|
for index in range(600))
|
|
blocks = ft.split_text(text, True)
|
|
self.assertGreater(len(blocks), 1)
|
|
for block in blocks:
|
|
for line in block.splitlines():
|
|
self.assertTrue(line.startswith("["))
|
|
|
|
def test_a_single_line_longer_than_the_limit_is_kept_whole(self):
|
|
text = "x" * (ft.CLEANUP_CHUNK_CHARS + 100)
|
|
self.assertEqual(ft.split_text(text, True), [text])
|
|
|
|
|
|
class SplitWav(DikteTest):
|
|
def wav(self, seconds, name="in.wav"):
|
|
return make_wav(self.path(name), silence(seconds))
|
|
|
|
def test_a_short_file_is_handed_back_as_it_is(self):
|
|
path = self.wav(2)
|
|
self.assertEqual(ft.split_wav(path, self.root), [(path, 0.0)])
|
|
|
|
def test_a_long_file_is_cut_at_the_chunk_length(self):
|
|
path = self.wav(5)
|
|
with mock.patch.object(ft, "CHUNK_SECONDS", 2):
|
|
chunks = ft.split_wav(path, self.root)
|
|
self.assertEqual([offset for _, offset in chunks], [0, 2, 4])
|
|
|
|
def test_the_chunks_add_up_to_the_original(self):
|
|
path = self.wav(5)
|
|
with mock.patch.object(ft, "CHUNK_SECONDS", 2):
|
|
chunks = ft.split_wav(path, self.root)
|
|
total = 0
|
|
for chunk_path, _ in chunks:
|
|
with contextlib.closing(wave.open(chunk_path, "rb")) as wav:
|
|
total += wav.getnframes()
|
|
self.assertEqual(wav.getframerate(), 16000)
|
|
self.assertEqual(total, 5 * 16000)
|
|
|
|
def test_the_chunks_do_not_write_over_each_other(self):
|
|
path = self.wav(5)
|
|
with mock.patch.object(ft, "CHUNK_SECONDS", 2):
|
|
chunks = ft.split_wav(path, self.root)
|
|
self.assertEqual(len({chunk for chunk, _ in chunks}), len(chunks))
|
|
|
|
|
|
class Transcriber(DikteTest):
|
|
"""The chain, with ffmpeg and both API calls faked."""
|
|
|
|
def setUp(self):
|
|
super().setUp()
|
|
self.source = make_wav(self.path("input.wav"), tone(1.0))
|
|
self.conf = self.config(openrouter_api_key="sk-or-test")
|
|
|
|
def run_chain(self, timestamps=False, cleanup=False, transcript="raw text",
|
|
segments=None, cleaned="clean text", fail=None):
|
|
worker = ft.FileTranscriber(self.conf)
|
|
done, failures, progress = [], [], []
|
|
worker.finished.connect(lambda *args: done.append(args))
|
|
worker.failed.connect(failures.append)
|
|
worker.progress.connect(progress.append)
|
|
|
|
def to_wav(path, workdir):
|
|
return make_wav(self.path("converted.wav"), tone(1.0))
|
|
|
|
with mock.patch.object(ft, "_to_wav", side_effect=to_wav), \
|
|
mock.patch.object(ft.shutil, "which", return_value="/usr/bin/ffmpeg"), \
|
|
mock.patch.object(api, "transcribe",
|
|
side_effect=fail or (lambda *a, **k: transcript)), \
|
|
mock.patch.object(api, "transcribe_segments",
|
|
return_value=segments or [(0.0, 1.0, "raw text")]), \
|
|
mock.patch.object(api, "cleanup", return_value=cleaned) as cleanup_call:
|
|
# The chain is run here rather than through start(): its signals are
|
|
# emitted from the worker thread, and a queued connection would need
|
|
# an event loop to deliver them. This is the same code, one frame down.
|
|
worker._work(self.source, timestamps, cleanup)
|
|
return done, failures, progress, cleanup_call
|
|
|
|
def test_plain_text_out(self):
|
|
done, failures, _, _ = self.run_chain()
|
|
self.assertEqual(failures, [])
|
|
self.assertEqual(done[0][0], "raw text")
|
|
|
|
def test_cleanup_replaces_the_text(self):
|
|
done, _, _, _ = self.run_chain(cleanup=True)
|
|
self.assertEqual(done[0][0], "clean text")
|
|
|
|
def test_cleanup_is_told_it_is_writing_subtitles(self):
|
|
_, _, _, cleanup_call = self.run_chain(cleanup=True)
|
|
prompt = cleanup_call.call_args.args[2]
|
|
self.assertEqual(prompt, self.conf.cleanup_prompt(subtitles=True))
|
|
|
|
def test_timestamps_come_back_as_segments_and_as_stamped_lines(self):
|
|
done, _, _, _ = self.run_chain(
|
|
timestamps=True, segments=[(0.0, 1.0, "one"), (2.0, 3.0, "two")])
|
|
text, segments = done[0]
|
|
self.assertEqual(text, "[00:00] one\n[00:02] two")
|
|
self.assertEqual(len(segments), 2)
|
|
|
|
def test_no_ffmpeg_installed(self):
|
|
worker = ft.FileTranscriber(self.conf)
|
|
failures = []
|
|
worker.failed.connect(failures.append)
|
|
with mock.patch.object(ft.shutil, "which", return_value=None):
|
|
worker._work(self.source, False, False)
|
|
self.assertIn("ffmpeg", failures[0])
|
|
|
|
def test_an_api_failure_is_reported_rather_than_raised(self):
|
|
def boom(*args, **kwargs):
|
|
raise api.ApiError("OpenAI rejected the API key")
|
|
_, failures, _, _ = self.run_chain(fail=boom)
|
|
self.assertIn("rejected", failures[0])
|
|
|
|
def test_empty_text_is_not_sent_to_cleanup(self):
|
|
_, _, _, cleanup_call = self.run_chain(cleanup=True, transcript="")
|
|
cleanup_call.assert_not_called()
|
|
|
|
def test_a_second_start_while_one_is_running_is_ignored(self):
|
|
worker = ft.FileTranscriber(self.conf)
|
|
worker._thread = mock.Mock(is_alive=lambda: True)
|
|
self.assertTrue(worker.busy)
|
|
worker.start(self.source, False, False)
|
|
self.assertTrue(worker.busy)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|