From Meeting Audio to English Subtitles and Dubbing
Résumé
Pixel’s announcement explores how LDX Voice transforms Japanese meeting recordings into English subtitles and dubbed audio using CastVoice, RefineLoop, and RenderVoice. It covers human review checks, fixing misheard terms, translation revisions, and whether the dub fits the original timing.
DEV Community
From a meeting recording to English subtitles and a dub: two human checks, then misheard terms fixed and the translation reviewed and revised
Earlier posts in this series: I turned our language AI API into an agent plugin, then ran the same 1-minute recording through three agents A non-engineer CEO tried to connect Copilot Studio to LDX hub. Here's what happened. A non-engineer CEO connected Copilot Studio to LDX hub. It took 8 errors and 3 hours. I skipped Power Automate and connected LDX hub directly via MCP. It worked in 2 hours. I batch-processed 20 meeting minutes with Power Automate + LDX hub. We fed meeting minutes to AI without structure, and lost 70% of the data we needed for decisions Building a Technical Literature Dashboard with Power Automate, Power Apps and LDX hub StructFlow Last week I ran the same 1-minute recording through Claude Cowork, Google Antigravity and OpenAI Codex CLI, and got the same subtitles from all three. This post is about what LDX Voice does in between: how a Japanese recording becomes English subtitles and a dubbed audio track. I'm not an engineer. I wanted answers to three questions: How many misheard names can be fixed before translation, and how often does a person have to step in? How many review passes does the translation need before it settles? Does the English dub fit the original timing? What LDX Voice is made of LDX Voice combines three LDX hub services. LDX hub service Role in LDX Voice CastVoice Transcription RefineLoop Translation, then review and revision RenderVoice Speech synthesis for dubbing (Lab edition) Input Output Audio (MP3, WAV; M4A and MP4 are converted to WAV) Source and translated SRT Teams or Zoom transcript (VTT) Translated SRT and VTT, with speaker names restored Video Video with subtitles (a subtitle track or burned in) Translated SRT (Lab edition) Dubbed audio (WAV) and subtitles timed to it With a VTT file there is no transcription step. Speaker names are not sent for translation. They go back in when the translated VTT is written. Stereo recordings are mixed down to mono before transcription, so a recording with one speaker on each side doesn't lose one of them. The flow Check 1: languages, models, glossary, whether to dub, and approval of the cost estimate. Transcribe with CastVoice into SRT. A Teams or Zoom VTT is converted to SRT instead, at no charge. Check 2: misheard terms and gaps in the audio. Normalize the transcript. Timecodes don't change. Translate and review with RefineLoop: pass 1 translates, passes 2 to 7 review and revise. Output: translated SRT and VTT, plus a comparison report. Optional: add the subtitles to the video. Optional (Lab edition): dub with RenderVoice, then fit each block to its time slot. A person answers exactly twice. Whether to dub is asked in Check 1, so it doesn't add a third check. After Check 2, everything runs to the final report. What the person looks at in the end is the translated SRT and the comparison report. Before any paid step, the cost estimate is shown and the job waits for a reply. The first request doesn't count as approval. Names: catch them before translation Speech recognition gets product names and people's names wrong. Translating the errors makes them worse. So the transcript is checked for candidates first: matches with glossary readings and with misrecognitions recorded before, homophones (in my case, my own surname came out as a different one with the same sound), words that make no sense in context, long gaps, and numbers cut in the middle. The candidates come in three groups. Group What How I answer A: spelling Product names, abbreviations, capitalization that context makes clear "All of A is OK" in one reply B: needs judgment People's names, homophones, words that make no sense OK or the correct spelling, item by item C: audio Missing speech, numbers cut across blocks Re-transcribe that segment, fill it in from a script, or leave it marked The same table asks for the English rendering of each name. If dubbing is on, it also asks how each term should be read aloud. Confirmed misrecognitions are recorded in the glossary, so the same ...

Powered by Readybot.io