The five events

These are delivery events where machine translation was used, suspected or disputed. Suspicion in a client complaint is not confirmation that machine translation was used.

The five events
WhenWorkWhat happenedOutcome
2021 (E1)Japanese-to-English subtitle sampleThe client rejected the sample and alleged that the translator had only edited machine-translated sentences. That allegation was not confirmed. The client said a large project was lost.The task payment was refused. Nothing later is recorded.
10 Jun-20 Jul 2022 (E2)Tamil subtitle MTPE, with one Telugu fileRepeated QMS failures led to urgent rework. Two files were redone by another team and not invoiced. On 15 Jul, three reworked files still failed a second QMS check. The client reported unread instructions and disturbed timestamps; the linguist's explanation about non-synced timings was not independently verified.On 20 Jul, the client said the next batch might not be assigned. No re-scored acceptance result is recorded.
Dec 2022-Jan 2023 (E3)Korean-to-Japanese subtitle test for a new end-client accountThe end client listed timing, unrelated words and suspected machine translation. The partner rejected the first reply as too general and demanded point-by-point answers, including whether raw machine translation had been used. Its use was not established.The partner said there would be no payment without a detailed answer that day. The test result and payment outcome are not recorded.
Jul 2026 (E4)Live-event SRT reviewMoniSa judged the supplied subtitles machine-generated and in need of heavy correction. The client did not dispute that assessment and said it would discuss reclassifying the work as MTPE.The actual reclassification is not recorded.
28 Jul-3 Aug 2026 (E5)Machine-subtitle review for short-form online video; videos of 2-5 minutes with a 60-90-minute turnaroundReviewer replies arrived late, some on later days, and one reviewer asked for missing scope. Reviewers cited time zones and prior bookings; a late file was reassigned.The record does not establish that every urgent file obtained a reviewer.

What we read into them

1. Observed: in E1, the client rejected the sample and alleged that the translator had only edited machine-translated sentences. In E3, suspected machine translation was one of the end client's complaints, alongside timing and unrelated words, and the partner then demanded to know whether raw machine translation had been used. In E4 the same question came up without a dispute: MoniSa judged the supplied subtitles machine-generated, and the client said it would discuss reclassifying the work as MTPE. Our reading: an unstated workflow per file can turn a quality discussion into a dispute about what was bought and whether it should be paid for.

2. Observed: E2 includes disturbed or non-synced timings, with different accounts of the cause; E3's complaint lists timing alongside unrelated words. Our reading: checking the words does not replace a separate timing and reading-speed pass after the edit.

3. Observed: on 15 Jul 2022, three reworked files failed a second QMS check. Our inference: this fits a repair that fixes flagged lines while leaving the same error class elsewhere. The records do not prove that this was how the rework was done.

4. Observed: E5 records late replies, missing scope and a reassigned file; the reviewers cited time zones and prior bookings. Our reading: for same-day review, reviewer availability was a constraint before editing could begin.

What the records cannot tell you

The client QMS attachments are absent, so we cannot say whether accuracy, timing or instruction errors drove the E2 failures. There are no edit-distance or editing-time measurements. No event records a pass after rework or a re-scored acceptance; in order, the five trails end at a refused payment, a warning that the next batch might not be assigned, a demand for answers, a discussion of reclassification and a reassigned file. The machine-translation allegations in E1 and E3 remain unconfirmed.

Five events across six years cannot supply a pass rate or failure rate. The archive is uneven: it holds five records for 2020-2021 and 49 for 2026. These events describe what is visible in this archive, not how often subtitle MTPE succeeds or fails.

What to write into your subtitle MTPE spec

These are our recommendations from the events, not additional findings in the records.

Name the workflow per file: human translation, full post-editing, light post-editing, or review of client-supplied machine output. Ask the vendor to state it again at delivery.

Give timing and reading speed their own check and signoff after the text edit. After a first failure, ask for the error class, a whole-file search for the same class and a second reviewer. Accept the re-scored result rather than treating redelivery itself as acceptance.

For urgent work, book a named reviewer for each time zone. Put the language, minutes, tool, review-versus-translation scope and due time in the first message.

Research that connects

CSA Research Jan 2023: TechStack: Automated Captioning and Subtitling Solutions, an analyst view, describes how errors multiply when transcription, captioning and machine translation are chained. That connects with the need to check timing as well as text.

Slator 2023: Slator Pro Guide: Subtitling and Captioning quotes one vendor executive, in a June 2023 interview, calling the human post-editing needed to bring machine translation for premium media to a usable level "onerous". In its section on automated live translated captions, it reports another vendor's view that clients who want a premium product remain wary of chaining speech recognition and machine translation, where a recognition error can carry into the translation. That connects with the difference between receiving output and accepting a checked result.

CSA Research for the Interpreting SAFE-AI Task Force Feb 2024: Perceptions on Automated Interpreting reports 42% poor versus 38% good for automated subtitling; adding machine translation lowered the rating. This is context from that research, not a measurement of our five events.

These research connections help interpret the questions raised here. They do not establish the causes of the events in our own records.

Method note

Source: MoniSa's media delivery archive contains 133 records dated 14 Jul 2020 to 15 Oct 2026, drawn from project, operations and information mailboxes and Drive QC logs. They were compiled into an internal ground-realities file on 21 Oct 2026 and read on 16 Nov 2026.

Selection: we selected every subtitle event in which machine translation was named, suspected or used by the client, MoniSa or the task. We counted an event rather than each record: several messages about the same event were grouped together. The five events draw on one, 13, five, one and seven records respectively; the long E2 chain covers roughly six or seven distinct incidents within that chain.

We separated observed facts, a participant's stated cause, our inference and the outcome actually recorded. Where there was no later message, we recorded the outcome as not recorded. We excluded a 2022 file with 60-70 changes over 3-4 hours because the record did not say machine translation was involved, and excluded voice-over, dubbing and trailer QC that were not MTPE. Client identities, people and commercial details are omitted. The archive's missing reports, uneven coverage and unconfirmed allegations limit what can be concluded.

The full report is being published as an open research record; its DOI link will be added here.

Plan the next step

Use these pages to continue the scope discussion.

Specify the test before editing

Use the recommendations above when writing the subtitle test brief.

  • State the workflow for each file and repeat it at delivery.
  • Separate text editing from timing and reading-speed signoff.
  • Require a re-scored result after rework.

Questions the record leaves open

These signals call for clarification rather than an assumed cause.

  • Suspected machine translation presented as confirmed use.
  • Redelivery presented as a recorded quality pass.
  • A flagged-line repair with no whole-file check for the same error class.

Scope a subtitle MTPE test

Put the review scope in the first message.

  • Language, minutes, tool and due time.
  • Whether the task is translation or review of supplied machine output.
  • A named reviewer and the timing check required after editing.