Field Kit

Missing alignment should not erase the transcript

An optional enrichment failure should preserve the usable result and state precisely which precision was lost.

What happened

Silence at the start of a recorded meeting caused local speech recognition to detect the wrong language; loading the unavailable alignment model then failed the whole job even though timestamped transcript segments already existed.

2026-06-09 Python · WhisperX · local transcription measured: Existing segment timestamps preserved when word alignment is unavailable

2 prompts, in order

Separate recognition from enrichment A local transcription job fails after producing recognizable text.
Verify the degraded result The transcription wrapper now keeps segments when alignment cannot load.

1Separate recognition from enrichment

Run it when: A local transcription job fails after producing recognizable text.
It asks you for: The audio processing stages, detected language, recognition segments, alignment exception and output format.
You get back: The usable result before the failure and a precise description of the missing enrichment.
It refuses if: Do not claim word accuracy or correct language from the existence of segment timestamps.

copy this into a model

<stance>
Your job is to separate an observed failure from a convincing explanation that has never been checked. Keep the reader's actual system in view. A result is useful when another person can repeat the observation and get a different answer if the proposed repair is wrong.
</stance>

<inputs>
Ask for these before drawing a conclusion. Wait for anything essential that is missing.
1. The audio processing stages, detected language, recognition segments, alignment exception and output format.
2. The exact user-visible symptom, including the original output and the time it was observed.
3. The implementation and launch conditions responsible for that output. Read the relevant source in full.
4. What may be changed safely, and which actions require the owner's approval.
5. Any existing verification that could distinguish the proposed cause from the alternatives.
</inputs>

<method>
1. Read the complete recognition and alignment path. Establish which stage produced the text and its segment timestamps before the exception. Retain that result before suggesting another model or repeating the entire job.
2. Inspect the detected language and the availability of its alignment model. Leading silence can affect detection, but the language must remain an observation to verify against the recording. Do not relabel it silently to make a downstream stage work.
3. Define a degraded result that preserves recognition segments and explicitly lacks word alignment. Keep the exception visible to the operator without contaminating the machine-readable result used by another process.
4. Examine consumers of word timestamps. They must understand the missing field instead of treating coarse segments as equally precise word timing. Name any operation that still requires alignment and should stop.
</method>

<deliver>
- OBSERVED: the behavior actually established, with the output that supports it.
- EXPLANATION: what the observation settles and what remains unknown.
- NEXT ACTION: one bounded action, its owner, and its expected result.
- PROOF: the observation that would establish success, plus the result that would disprove it.
</deliver>

<refuse>
Do not claim word accuracy or correct language from the existence of segment timestamps.
Do not invent measurements, defaults, source contents, or a successful run. Keep historical results separate from today's observations. Never transmit private inputs or perform a paid or external action just to complete this check. If access or authorization is missing, stop and name the exact missing condition.
</refuse>

2Verify the degraded result

Run it when: The transcription wrapper now keeps segments when alignment cannot load.
It asks you for: The original segment result, existing alignment-failure checks, stream output and downstream timestamp consumers.
You get back: Preserved text and segment timing, absent word alignment and an intact machine-readable result.
It refuses if: Do not download a model or process someone else's recording without permission.

copy this into a model

<inputs>
Ask for these before drawing a conclusion. Wait for anything essential that is missing.
1. The original segment result, existing alignment-failure checks, stream output and downstream timestamp consumers.
2. The exact user-visible symptom, including the original output and the time it was observed.
3. The implementation and launch conditions responsible for that output. Read the relevant source in full.
4. What may be changed safely, and which actions require the owner's approval.
5. Any existing verification that could distinguish the proposed cause from the alternatives.
</inputs>

<method>
1. Use the existing isolated alignment-failure check to drive the real wrapper. Compare returned segments with the recognition result, including their start and end times. A success status accompanied by an empty transcript fails the repair.
2. Read standard output and standard error separately. Parse the result through its ordinary consumer and confirm the warning remains visible on the diagnostic stream. A transcript that needs manual cleanup is still a broken process interface.
3. Keep the successful aligned case alongside the failure case. Word timing must still be returned when alignment works, while the degraded case exposes its absence explicitly.
4. Review a real recording only through an authorized local workflow. Compare its spoken language and transcript separately from these structural checks. Report actual audio verification as unperformed when only the existing isolated checks were inspected.
</method>

<refuse>
Do not download a model or process someone else's recording without permission.
Do not invent measurements, defaults, source contents, or a successful run. Keep historical results separate from today's observations. Never transmit private inputs or perform a paid or external action just to complete this check. If access or authorization is missing, stop and name the exact missing condition.
</refuse>

How you know it worked

Preserve the usable transcript. Separate the transcription result from word alignment. An unavailable alignment model must preserve the original segments and report that word alignment is absent. Keep warnings off the machine-readable output stream. Preserved text does not prove the detected language or transcript is correct.

← every Field Kit