Does video translation and lip sync need the cloud? For creators who want footage to stay on their own computer, the practical question is how long local processing takes.
We tested Bonframe on a Mac with an M5 Pro, 15-core CPU, 16-core GPU and 24GB of unified memory. A 46.16-second, two-speaker video took 4 minutes 34 seconds for SD Fast lip sync, or 6 minutes 38 seconds with the HD model in Fast mode. No remote GPU performed the lip-sync inference.
Those are useful local turnaround times. HeyGen's published processing guidance gives them context, provided we keep the workflows and measurement boundaries distinct.
Watch the result: Chinese-to-English video translation and lip sync
This previously completed Chinese-to-English video shows the finished output. Play it to inspect the English voice track, mouth movements and speaker changes.
The English video is a result demonstration. The timing measurements below use the earlier Chinese-to-Japanese test runs; they are not generation times for the embedded English video.
What HeyGen's five- and ten-minute figures mean
HeyGen's help center gives different estimates for different workflows: Video Translation, including Proofreading, takes about five processing minutes per minute of content; Avatar Studio usually takes about ten, and can be faster. Resolution, engines, subscription tier and submission time can affect the result. Source: HeyGen Video Processing Times
For a video-translation discussion, the relevant published reference is approximately five times the source duration. Combining it with Avatar Studio's ten-minute estimate would mix two tasks. These ratios describe waiting time: a higher number means a longer wait, not faster-than-real-time generation.
There is also a measurement difference. Our Bonframe timings cover the lip-sync stage after translation and speech generation. HeyGen's figure describes video translation. We did not submit the same clip to HeyGen, and the local numbers exclude ASR, translation and TTS. This is a comparison of processing-time scales, not an equal-workload speed ranking.
MacBook Pro-class hardware, measured on a Mac mini
The tested configuration was M5 Pro with a 15-core CPU, 16-core GPU and 24GB of unified memory, without an external GPU. Apple also lists this chip and memory configuration for the 14-inch MacBook Pro.
The actual test machine was a Mac mini. It was not a MacBook Pro benchmark. The shared specification makes the result relevant to owners of similarly equipped laptops, but cooling, power mode and sustained load still require testing on the laptop itself.
The source was a 46.16-second, 1440×1080, 25fps video with two speakers and speaker changes. Each configuration reused the same Japanese translation and voice track. Every result retained the 46.16-second timeline and 1,154 output frames.
Timing included lip-sync planning, cropping, model loading, inference, decoding and composition. It did not isolate only the fastest GPU operation, and it did not include a separate final-file export.
Local lip-sync results
The current-version records were:
- SD Fast: 4m 34.12s, or 5.94× the source duration.
- SD Lite: 6m 52.66s, or 8.94×.
- HD Fast: 6m 37.96s, or 8.62×.
- HD Lite: 8m 44.41s, or 11.36×.
- HD Pro: 11m 49.05s, or 15.36×.
Each figure is one measured run of this clip. SD and HD identify Bonframe model tiers; Fast, Lite and Pro are its encoding options. They do not establish equal quality with HeyGen's Speed or Precision engines. HeyGen describes those choices in its video-translation guide.
The higher-quality options took longer. The Fast result should not stand in for every configuration. SD Pro was excluded because its available record predates the current decoding optimization.
How close is the fastest local result?
Applying HeyGen's published five-times-duration reference to 46.16 seconds gives 3m 50.8s. That is an illustrative calculation, not an actual HeyGen run or a promised completion time.
Bonframe's SD Fast lip-sync task took 4m 34.12s: approximately 43 seconds, or 19%, longer than that reference. A 16-core-GPU personal Mac can therefore perform this local task within a similar several-minute waiting scale.
The narrower conclusion matters: local lip sync on this M5 Pro is already practical enough to evaluate alongside cloud workflows. The measurements do not establish that the full Bonframe translation pipeline matches HeyGen's speed, or that both deliver the same visual quality. That would require the same source, target language, comparable output requirements and submission-to-result timing on both products.
What keeping the work on your Mac changes
For creators refining names, pronunciation and speaker assignments, local processing keeps the next lip-sync pass close to the existing footage and voice track. This test reused completed translation and TTS, then generated the different lip-sync configurations.
That workflow avoids uploading the full source to a remote GPU service for lip-sync inference. Jobs depend on your own hardware and current workload rather than a shared cloud processing slot. Local jobs still compete for the computer's resources; these single-task measurements do not predict simultaneous-job throughput.
For an owner of a similarly specified M5 Pro MacBook Pro, the desktop result is a useful starting point for testing an existing machine. It is not a guarantee of the same laptop runtime. See the Mac video-translation guide for setup and review steps, or the Windows GPU benchmark for a separate hardware study.
On this two-speaker clip, local SD Fast lip sync took about four and a half minutes; HD Fast took about six and a half. Creators who prefer local media processing now have a measured option to assess on their own footage, with explicit choices between speed and image quality.
Download Bonframe for your supported Mac and start with a short clip you are authorized to translate. Review the words, voice and picture before publishing the language version.
Measured and official sources checked November 5, 2026. One run per configuration, without a system file-cache flush; no long-video scaling or repeated-run average is claimed. Timings are local lip-sync task wall times, not full translation-pipeline or export times. Outputs used the watermarked-preview path. The YouTube example demonstrates Chinese-to-English output; the timing sample used Japanese dubbing. No same-input HeyGen run or cross-product quality test was performed.