Transcript alignment for ASR: text correctness, clocks and boundary review Separate transcription from forced alignment, diagnose offset versus drift, and review word boundaries against a declared audio timeline. Transcript alignment estimates where supplied text occurs in audio. It is separate from deciding which words were actually spoken. A well-formed timestamp file can still align an incomplete transcript or use the wrong recording clock. A reliable workflow reviews the text, declares source and channel coordinates, generates candidate boundaries and checks representative timing failures before treating the result as a reference. Three jobs that should not be collapsed Transcription determines the words under a documented convention. Forced alignment locates supplied words using a compatible model and pronunciation resources. Boundary adjudication checks whether proposed intervals are suitable for the downstream task. Combining the jobs into one “aligned transcript” label hides where mistakes entered. The [Montreal Forced Aligner workflow](https://montreal-forced-aligner.readthedocs.io/en/v3.4.3/user_guide/workflows/alignment.html) documents a corpus, pronunciation dictionary and acoustic-model workflow. Its output is a tool result, not independent proof that a transcript is complete or that every boundary meets a project’s tolerance. Review the words before optimizing timestamps Listen for omitted short phrases, corrections, names, overlapping speech and clipped word starts. An aligner cannot label a word absent from its supplied text as if it had been transcribed correctly. It may distribute nearby boundaries in a way that looks plausible while the reference remains incomplete. Keep transcript revisions separate from timing revisions. For an invented example, “send it Thursday” supplied for audio saying “send it Friday—sorry, Thursday” has a missing correction. Fixing the word list changes the target; moving the timestamp of Thursday does not resolve that omission. Preserve the old version and the reason for the edit so evaluation changes remain explainable. Declare the clock and channel Record the source file identifier, selected channel, offset from the source session and time units. If a clip begins 120 seconds into a session, a clip-local word at 2.4 seconds maps to session time 122.4 seconds only when the declared transformation is a simple offset. Cuts, speed changes and alternate edits require a more complete mapping. Include the exact audio derivative used by the aligner and its processing history. A timestamp reference generated on an enhanced or trimmed edition should not silently accompany a different original file. Hashes help identify files, while a lineage record explains the relationship; neither substitutes for boundary inspection. Worked diagnostic: offset is different from drift These invented checkpoints compare reviewed event times on one source with the corresponding times in an annotation file. The annotation starts 0.2 seconds late and the difference grows. A single constant correction cannot explain all three points. Checkpoint Reviewed source time Annotation time Difference Start phrase 10.0 s 10.2 s +0.2 s Middle phrase 110.0 s 110.7 s +0.7 s Late phrase 210.0 s 211.2 s +1.2 s Interpret the pattern before applying a correction In this deliberately linear example, the difference grows by 0.5 seconds over each 100-second interval. That suggests checking clock scale, resampling metadata or edition changes. It does not identify the cause from timestamps alone. Inspect more anchors and the actual waveform before applying an affine correction. An abrupt jump would suggest a different investigation, such as an inserted advertisement or removed pause. A problem confined to one speaker may indicate channel or overlap handling. Keep a residual-error report after any correction; a transformation that fixes two anchors can still fail between them. Define review criteria around the downstream use A subtitle display, word-level search index and turn-taking evaluation can tolerate different kinds of boundary error. Define which boundary is judged, the reference convention, treatment of silence and overlap, and how uncertainty is represented. Do not import a universal millisecond tolerance from an unrelated task. Sample difficult conditions deliberately: rapid speech, rare pronunciations, music beds, multiple speakers and long pauses. Report absolute start and end deviations against adjudicated anchors and retain example-level failures. A median alone can hide a small set of unusable clips. Validate structure and linguistic resources Check that intervals are ordered, remain within the declared file duration and have valid source identifiers. Decide whether zero-length entries are permitted and how unintelligible speech is encoded. Use language and pronunciation resources compatible with the audio rather than assuming one dictionary covers every accent or code-switched word. MFA’s [corpus validation documentation](https://montreal-forced-aligner.readthedocs.io/en/v3.4.3/user_guide/workflows/validate.html) is a useful preparation reference. Validation findings still require interpretation: a structurally valid corpus can contain a wrong transcript or use a source edition that differs from the customer’s copy. Ask for a reviewable alignment package Specify source and derivative identifiers, transcript policy/version, timing convention, channel map, alignment-tool configuration, reviewed sample report and unresolved exceptions. Separate automatically proposed boundaries from human-reviewed ones in the delivery description. AIPodcast can discuss whether a proposed source supports the required preparation and review. This article does not promise word-aligned inventory, a particular accuracy level or a fixed alignment service. Agree those requirements against the selected material and inspect a representative sample. Does forced alignment correct the transcript? It locates supplied text under a model. Missing or incorrect words need transcription review; plausible boundaries do not make the text correct. Can I shift all timestamps by one offset? Only if representative anchors support a constant-offset explanation. Drift, edits and speed changes require different mappings and residual checks. Must every ASR training example have word boundaries? No. The selected model recipe determines the required targets. Add word-level timing when the intended workflow needs it, and specify its review status separately. Research method: Fiund research desk; public sources, explicit worked-example labels and project-specific availability review. Editorial contact: hello@fiund.com. Sources MFA 3.4.3 alignment workflow: https://montreal-forced-aligner.readthedocs.io/en/v3.4.3/user_guide/workflows/alignment.html MFA corpus validation and first steps: https://montreal-forced-aligner.readthedocs.io/en/v3.3.4/first_steps/