Bilingual Subtitles Without Uploading a Single Frame of Video

You have a video and you need subtitles — ideally in two languages. The commercial route is a transcription service that bills by the minute, holds your audio under its own retention policy, and usually caps the free tier just below the length of anything useful. For lecture recordings, internal training, interviews, and anything confidential, that route is closed before you start.

The entire pipeline runs locally. Here is how to do it so the result is actually good, because “it produced a transcript” and “the subtitles are usable” are different bars.

Step 1: transcribe locally

Load the audio or video into the speech-to-text tool and let Whisper run on your machine. Practical notes that change the result:

• Set the language explicitly if you know it. Auto-detection is good, but a hint removes an entire class of mis-detections on short or noisy clips.
• Pick a model size deliberately. The smaller tiers are fast and fine for clear single-speaker audio; the mid tier is worth the extra download for accented speech, technical vocabulary and names.
• Denoise first if the recording is bad. A noisy room hurts recognition more than any other single factor, and denoising also runs locally.

The model downloads once and is cached. After that you can finish the job with the network unplugged, which is sometimes the actual requirement.

Step 2: review with timestamps, not with a wall of text

Whisper is strong and it is not perfect. Budget five minutes of correction per hour of audio and focus on the errors that matter:

• Proper nouns — speaker names, product names, place names. Transcribers render the unknown phonetically, and these are the errors viewers actually notice.
• Homophones in technical speech — the difference between a correct and an embarrassing subtitle is often one word.
• Sentence boundaries — the model has to guess where thoughts end, and interrupted speech makes that guess unreliable.

Reviewing against the timeline rather than as plain text is much faster, because you can scrub to a suspicious line and hear it instead of inferring it.

Step 3: produce the second language

With a clean transcript, translation is far more reliable than translating raw model output — you are no longer asking anything to translate a misheard word. Two approaches:

• Translate the subtitle file — preserves your timing and your corrections. This is usually the right choice.
• Translate first, then re-time — occasionally better for languages whose sentence length differs greatly from the source, where a one-to-one mapping produces unreadably long lines.

Either way, have a human read the result once. Machine translation of technical or legal content is a draft, and a subtitle is unforgiving: it is on screen for two seconds and there is no footnote.

Step 4: export dual-language SRT

Export in the format your destination wants — SRT is the safe universal choice, VTT for web players. For bilingual display, the conventional layout is the original line on top and the translation below, in the same cue.

Two rules that separate readable subtitles from technically-valid ones:

• Keep lines short. Roughly two lines, around 40 characters each. Viewers read subtitles; they do not study them.
• Respect shot changes and pauses. Splitting a cue across a cut feels wrong even when the timing is technically fine.

Why local matters for this particular job

Subtitling work disproportionately involves material that was never meant to leave the building: internal all-hands, customer support calls, medical and legal consultations, interviews conducted under an agreement, lecture recordings containing student questions. Every minute of that audio is identifiable speech with names in it.

Local transcription changes the cost structure as well as the privacy one. There is no per-minute fee, so a three-hour recording costs the same as a three-minute one — which means you can subtitle things that would never have justified the invoice.

Start with the speech-to-text tool, then move to the subtitle studio for timing and export. The first time you finish a full dual-language subtitle file without a single byte leaving your machine, the cloud workflow starts to look like an odd thing to have normalised.