Translation Test Methods and Cases

Scope

Review two VoxSail translation cases, their meaning checks and workflow timing estimates, with clear limits on what the results can establish.

These cases come from the test environment, not independent third-party certification. They do not alone substantiate the homepage figures of 98% meaning retention or 10x faster localization across all videos and languages.

How Meaning Is Reviewed

VoxSail first transcribed each source. A user checked it against the original audio, then the reference and semantic groups were frozen before target-language generation. Each group must preserve its agreed facts, actors, quantities, negation, conditions and causal relationships. Natural paraphrasing is allowed; group counts are not changed after seeing the output.

Target text, final spoken meaning and end-to-end results are checked separately. Omissions, mistranslations and unsupported additions count as issues. Pronunciation, naturalness, noise and synchronization are separate dimensions. Subtitle cues are neither grammatical sentence counts nor the semantic scoring denominator.

Mandarin to English Case

The source is China News Service's video titled Why Do We Explore Mars? in English translation, about 3 min 18 sec long, with 38 source reference cues and 17 predefined semantic groups. The reference began as a product transcript; one user-confirmed correction was made before freezing it.

• The first target-language delivery used the Friendly Women / S2.1 Pro preset voice.
• On 2026-10-29, the user confirmed reviewing the complete English text and final dubbing, with no semantic omissions, mistranslations or unsupported additions found.
• Target text, final spoken meaning and end-to-end results each passed 17/17 groups.

English to Mandarin Case

The source is David Blake (Degreed)'s Why Open Education Matters, about 2 min 27 sec long, with 31 source reference cues and 12 predefined semantic groups. The English reference retains the source wording confirmed by the user.

• The first target-language delivery used the 王琨声音模型10.17t1 / S2.1 Pro preset voice.
• On 2026-10-30, the user confirmed reviewing the complete Chinese text, final dubbing and all 12 groups, with no semantic omissions, mistranslations or unsupported additions found.
• Target text, final spoken meaning and end-to-end results each passed 12/12 groups.
• The run merged reference cues 15 and 16, exporting 30 cues. The concatenated English text is unchanged, so the original 12-group rubric remains in effect; fewer cues are not automatically a semantic omission.

Interpreting 98% and These Results

The group pass proportion is passed groups divided by the predefined total for a selected video. It is not word error rate, voice similarity, accuracy across all languages or user satisfaction.

Both cases rely on one reviewer, a user and product operator. Group results were mapped from one explicit whole-video confirmation per case, not independently collected per-group judgments or second-reviewer blind review. References began as product transcripts, and videos were selected rather than randomly sampled. No reported issue does not guarantee that no undetected error exists.

The 17/17 and 12/12 results support these specific cases, not 29 independent video samples or a general 98% meaning-retention validation. Broader substantiation requires predefined metrics and sampling, more independent videos across language directions, topics and audio conditions, retained failures and disputes, and independent bilingual review.

Workflow Timing Method

The manual baseline includes listening transcription, source correction and segmentation, manual translation, translation revision, human voice recording with trimming and synchronization, subtitle layout and final delivery. It is not CapCut ASR or machine translation followed by corrections. Actual user-supplied stage records total 134 minutes for Mars and 125 minutes for Open Education. The final stage includes rendering, so these are not pure active labor totals.

VoxSail project pages show 5 min 30 sec and 4 min 15 sec from creation to completion. The user also reported approximately up to two minutes of operating time per video, excluding playback, with review at 2x speed, upload taking a little over ten seconds, first-use subtitle styling taking about 30 seconds and download taking about 10 seconds. Rendering is already included in processing according to the user and is not added again.

These are case calculations combining system records and approximate user reports. Playback is derived from output duration divided by two, not stopwatch timing; operations and playback may overlap. Each video is treated as a separate first-use styling scenario, not evidence a shared template was actually recreated twice. Template reuse is not assumed to require zero operating time.

• Mars: adding these components as a first-use scenario gives about 10 minutes; 134 minutes divided by that approximate value is about 13x.
• Open Education: the equivalent scenario gives about 8 min 20 sec; 125 minutes divided by that approximate value is about 15x.

Interpreting 10x and Timing Limits

The approximate 13x and 15x ratios are consistent with a 10x order of improvement, but are not controlled matched end-to-end measurements. Both historical runs had embedded subtitles disabled; a reported styling allowance does not establish timed captioned deliveries for those outputs. Human recording and preset TTS also use different production methods, without proving equal voice identity or artistic performance.

A full efficiency validation needs unified start/end timing, matched deliverables and explicit quality criteria, including necessary review, rework and export. Approximate inputs must not be relabeled as precise measurements. Two cases do not establish that all tasks are 10x faster; network conditions, duration, audio quality, models, load and human revisions can affect results.

Use of Results and Media

This page provides only method, condition and result summaries. Audio and video remain private; no internal task or account information is provided. Cases use first deliveries from the test environment on 2026-10-29 and subsequent user confirmations, not all product versions or production-environment results.

Source videos are translation subjects, not endorsements by their creators, and their viewpoints are not endorsed facts. Semantic acceptance does not grant voice-showcase rights. Any public media use requires a separate check of the source version, attribution, adaptation notices, third-party content and relevant voice permissions.