Instagram transcript translator: transcribe first, translate second
They are two steps, and this tool only does the first one. Ask it for words and you get the words that were spoken — a Reel in French comes back in French, never in English — so translating means taking that text somewhere else afterwards. Doing it in that order is the steadier route for a concrete reason: the transcript is a file you can read and check against the audio before anything is translated, and a wrong word you can see is a wrong word you can fix, while a wrong word inside a translation is invisible because the sentence still reads smoothly. The ordered route also survives the failure this model cannot avoid — one language is chosen for the whole file and it does not detect switches — because a mixed Reel whose second half came back wrong is caught at step one, before the translation has baked the mistake into a second language.
The rest of this page is the evidence: the two token ids in the speech model’s own configuration that decide transcribe versus translate, and why one of them can never be selected here; the three exported files and which of them survives a paste into a translation tool; what a second pass costs in text bytes when the transcript is 3,840,000 bytes of audio for a minute, and why the script is the cheapest input a translator can be handed; and the measured case where 19 of 338 words came back wrong, which is the check that has to happen before step two.
What the model’s own configuration says about translating inside it
Whisper does carry a translate task. Its configuration lists two task tokens and gives them ids, and the difference between them is one integer: task_to_id is {"transcribe": 50359, "translate": 50358} in both checkpoints this page uses. So nothing about the model prevents a translation; the question is whether the code path this page takes ever asks for one.
- The task slot is pinned to 50359 before any audio is read. The configuration’s
forced_decoder_idsis[[1, null], [2, 50359]]in bothonnx-community/whisper-baseandonnx-community/whisper-tiny. Position 1 is left null for the language token, position 2 is fixed at 50359 — the transcribe id — so the decode begins with transcribe already decided, and 50358 is 1 integer away and unreachable. - The argument app.js passes is the same constant. The call sets
task: 'transcribe', so the request cannot differ from the configuration. To get a translation out of this model the code would have to pass a different task, and the only place in the 22,349 bytes of app.js where the word translate appears at all is the library’s own documentation, not code that runs. - Neither the panel nor the picker offers a second language. The picker is filled from
lang_to_id, which holds 99 entries from id 50259 to 50357 — English at 50259, Chinese at 50260, Spanish at 50262. Every one of those is a language to listen to, not a language to write in; there is no target-language control to set, because a translator needs one and a transcriber does not. - The 99 ids end at 50357 for a reason. 50358 is translate, 50359 is transcribe, 50361 is the previous-timestamp token and 50363 is the no-timestamps token. The language block is closed at 50357 and everything above it is bookkeeping, which is why adding a language cannot silently swallow the task slot.
forced_decoder_ids [[1, null], [2, 50359]] writes the transcribe id. The translate task is not missing from the model — task_to_id gives it 50358, exactly 1 below transcribe — it is simply the id nothing here asks for. The 99 languages from 50259 to 50357 sit below both, because a language token says what the voice is, not what the output should be.Three files, and which one survives a paste into a translator
The export is a dead file, which is exactly what a second step needs: nothing about it phones home, and every translation tool accepts text. What differs is how much of the file is not words.
- The plain .txt is the cleanest input. It is the transcript body and nothing else — measured on a 161.2-second run, 1,875 characters for 338 words, with the timestamps switched off. That is the version to paste, because none of its 1,875 characters is bookkeeping that a translator would try to translate.
- With timestamps on it is 2,280 characters, and 405 of them are stamps. 27 rows × 15 characters each — a bracket, a 12-character timecode and a closing bracket — so 17.8% of the text is content a translator has to either keep verbatim or be told to ignore. The 12-character stamp itself cannot be mistranslated, but the digits and colons around it can be pulled into the surrounding sentence.
- The .srt and .vtt carry 648 characters of timing on top. Each of the 27 cues opens with an index line and two 12-character timecodes joined by an arrow, so 24 of those characters per cue are pure structure. A subtitle file is the right thing to hand a video editor and the wrong thing to hand a text translator, and the tool will produce both from the same run.
- Nothing in the export is translatable per line without a key. A cue holds three fields — a start, an end and the text — and none of them is a language or a speaker, so after translation the file no longer lines up with the audio unless the timecodes were left alone.
What the second pass costs, and why the script is the cheap end of it
The speech side of the run is expensive in bytes and the text side is not, which is the whole argument for doing the cheap step twice rather than the expensive one. The transcript is what you carry across; the audio does not go with it.
- One minute of speech is 3,840,000 bytes of audio for roughly 358 characters. At the measured rate of 1194 words in 600.3 seconds, a 30-second window holds 59.67 words; at 6 characters a word that is 358 characters, and 3,840,000 ÷ 358 is 10,727, so the translator is handed about one ten-thousandth of the byte count the recogniser had to read.
- A whole Reel is smaller than one window of audio the recogniser processed. A 120-second Reel is 6 windows of 480,000 samples each; its transcript at about 119 words per window is roughly 714 words and 4,286 characters, which in UTF-8 is 4,286 bytes against 7,680,000 bytes of 16 kHz mono audio.
- Translating does not re-run recognition. The 161.2-second run took 71.4 seconds of recognition, and the two long runs give 0.443 and 0.419 seconds per second of audio; re-running it because a translation needs a different cut of the audio costs that again, and the translation itself is text on a few thousand bytes. That is why the order is transcribe once then translate, not translate while transcribing.
- The download is already paid on the first file. 79.7 MB for the better-accuracy checkpoint and 43.6 MB for the smaller one, fetched once and then cached; a translation pass does not touch either. The 1,875-character transcript above is 0.0043% of the 43.6 MB checkpoint, and 3,840,000 bytes of mono float32 render for a minute of audio is 9.40% of it — the audio you decode is the item worth not repeating.
Why the order matters more when the audio is mixed
One language is chosen for the whole file. The picker holds 99 ids and one of them is applied to every 30-second window; the model does not detect a switch, and nothing in the pipeline compares window to window. A Reel that changes language half-way is therefore the case where a translation built on top of a bad transcript is worst: the wrong half comes back as confident nonsense in the source language, and the translation renders that nonsense fluently in the target language, where you can no longer tell it from a correct sentence.
- The measured version of that failure is 404 words where 54 were spoken. The same English clip run with the picker set to Spanish returned 404 words of repeated Spanish instead of the 54 English words it contains. If that had been translated, you would have been handed a fluent paragraph in a third language with no trace of where the error started.
- A language change inside one window cannot be labelled in the export. A cue carries a start, an end and text, and no language field, so even a correct mixed transcript loses the information of which line is in which language — and a translator reading the .txt has to guess.
- The honest fix is to split the file at the switch, not to translate through it. Six windows for a 120-second Reel, one language each; cut at the switch and run each half with the right language selected, and the two transcripts can then be translated as two separate texts. It costs a second recognition pass over one half, and it is the only route here that produces a transcript a translation can trust.
- A wrong transcript is worth checking before the translation, and it is checkable. Measured on a 338-word English run, 19 words came back wrong — 5.62% — across 11 places, 4 of them names. Reading the source transcript against the audio catches those 5.62%; reading the translation cannot, because the same sentence in another language reads perfectly well with the name still wrong.
What to do with the result, and what the tool will not do
Three shapes of answer, and they differ in where the second step happens rather than in whether it exists.
- A Reel you did not understand: transcribe with the spoken language selected, read the transcript, then paste the plain .txt into whichever translator you already trust. The 1,875 characters cost nothing to move and lose nothing in the move.
- A Reel you want to publish in another language: keep the timecodes. Translate cue by cue with the stamps preserved, then rebuild the .srt or .vtt from the translated lines, because the 12-character stamp is a fixed-width field and survives a round trip; a stamp re-typed by hand does not.
- A mixed-language Reel: split at the switch first, transcribe each half with its own language, translate each half separately, and join them at the cut. Doing it the other way round means translating a line whose source language you never confirmed.
- What is not here: no target-language control, no bilingual side-by-side view, no glossary, and no connection to any translation service — the page makes 2 network requests for the library and the language list, then 1 more class of request for the model weights, and sends 0 characters of your transcript anywhere.
Can I translate an Instagram Reel without transcribing it first?
Not with this tool, and the reason is in the model rather than in the page. Its configuration pins the task at position 2 of the decoder to the transcribe id, and the translate id one integer below it is never selected, so every run returns the spoken language. A tool that offers translation in one step is doing the same two operations behind one button — it still has to recognise the speech before it can produce a sentence in another language, and it still cannot tell you which source word it got wrong.
Does the transcript come back in the language I selected?
Yes, and that is a limitation as much as a feature. The picker’s 99 ids are languages to listen for — English at 50259, Spanish at 50262 — not languages to write in, so a French Reel transcribed with French selected returns French words. There is no setting that would return English instead, which is why the translation is a separate step and why the wrong picker choice damages the transcript before any translation exists.
Can I translate the subtitle file instead of the plain text?
You can, but keep the timing lines out of the translator’s hands. A 27-cue .srt carries 648 characters of timing — an index line and two 12-character stamps per cue, joined by an arrow — and a translator handed the whole file will try to make sense of those lines along with the words. Translating only the text line of each cue and leaving the stamps alone preserves both the sync and the fixed 12-character field.
What about a Reel that has two languages in it?
Transcribe it first and look for the switch, because the export will not mark it. One language is set for the whole file, applied to every 30-second window without comparison between windows, and a cue carries a start, an end and text with no language field. A 120-second Reel is 6 windows; cutting at the switch and running each half with its own language costs one extra pass over one half and is the only way to get two transcripts a translation can rely on.
Why does the order matter if the translation reads fine?
Because “reads fine” is the failure mode. Measured on a 338-word English run, 19 words came back wrong — 5.62% — and 4 of the 11 places were names, which is the class of error a reader cannot repair from context. In the source language you can hear the difference; once those 19 words are translated into a sentence that is grammatical in the target language, nothing on the page says which of them was never spoken.
Last checked: 9 November 2026. The two token ids, task_to_id at {"transcribe": 50359, "translate": 50358} and forced_decoder_ids [[1, null], [2, 50359]], together with lang_to_id holding 99 ids from 50259 to 50357 and no_timestamps_token_id at 50363, were read from generation_config.json fetched from huggingface.co for onnx-community/whisper-base (3,832 bytes) and onnx-community/whisper-tiny (3,772 bytes) on this machine today; both files agree field for field. That no translate path is taken comes from app.js, which passes task: 'transcribe' and passes nothing else that could override it. The byte figures are arithmetic on constants in app.js and on the Web Audio spec — RATE = 16000, one channel, a 32-bit float sample, so 64,000 bytes a second: 60 seconds is 3,840,000 bytes and 120 seconds is 7,680,000. The 1,875 and 2,280 characters, the 405 stamp characters, the 648 timing characters, the 338 words, the 27 cues and the 19 wrong words in 11 places (4 names, 5.62%) come from the 161.2-second run recorded on this machine on 27 October 2026. The 404-word and 54-word counts are the same measurement record’s picker test. The 1194 words in 600.3 seconds is a run on this machine on 27 October; the resulting 59.67 words per 30-second window and the 358-character figure that follows are arithmetic on it and assume 6 characters a word, which is an assumption about English prose rather than a measurement. 6 windows for 120 seconds is arithmetic on the pipeline’s window of 480,000 samples and its jump of 320,000. The 79.7 MB and 43.6 MB model sizes are the decimal sizes this site states elsewhere, and 3,840,000 ÷ 40,844,231 = 9.40%, with 1,875 ÷ 43,600,000 = 0.0043%, are computed from the tiny checkpoint’s measured byte count. The two network requests are the dynamic import of the library and the fetch of generation_config.json, both in app.js.
If the trouble is the language you picked rather than the translation, picking Hindi and other non-English languages covers how one token is placed at the front of every window. If the Reel switches language part-way, the mixed-language page works out where the switch falls inside six windows. If you want the translated lines back in a file, the SRT and VTT page covers the 12-character stamp, and the repurposing page covers what else the same transcript can feed.