This guide uses only the published package description, consulted 2026-11-08. It maps a workflow, not Python modules, dictionaries or an API signature.
| Familiar step | OneSpeech equivalent |
|---|---|
| ASR transcription | A SpeechRecognizer, or Speech.Transcribe for the integrated file path |
| Forced alignment for word timestamps | Aligner / Speech.Align, with nullable timing and per-word TimingMethod |
| Speaker diarization | A Diarizer producing original-clock Turn values |
| Assigning words to speakers | TranscriptDiarization.Merge, greatest covered duration with documented overlap rules |
| Saving segments/words | Immutable Transcript and schema-v1 JSON |
The integrated helper performs the supported workflow, but its defaults are different. The default diarizer does not detect overlap; use the explicit Nemotron upgrade if you need overlapping turns. An authorized local segmentation-3.0 snapshot can also supply overlapping turns with our clustering; community-1 has no native pipeline adapter. No noncommercial alignment weights are loaded.
Timing follows the supported timing policy, not a language-to-aligner dictionary copied from another project. Standalone alignment currently has a 60-second limit. Words without valid timing stay untimed, and words with no covering turn are Unassigned; neither is silently fabricated. Inspect timing and diarization results before treating labels or boundaries as ground truth.