An English-only transcription model does not switch languages when an Instagram Reel does. It maps every foreign word and every unfamiliar place name onto the closest English sound sequence it has learned, and the mistakes it makes are specific, countable and reproducible. In one freshly recorded clip, 334 spoken words came back as 326 tokens, with 28 substitutions, 22 deletions and 14 insertions, for 80.84% word accuracy.
Most pages answer with a list of tips. A list cannot tell you what a language switch costs, because that cost belongs to one language meeting one particular model on one particular recording. This page builds one clip that deliberately mixes two hard conditions, runs it through the exact model this site ships in the browser, and reports the word-level result with every input printed below.
What was put into the clip on purpose
Two conditions were built into the same recording. The first is non-English words spoken inside an otherwise English sentence: bonjour, merci, au revoir, gracias, ciao, danke, sayonara, arigato, adios, buenos noches, schnitzel, gelato, ristorante, rendezvous and amigo all appear in the exact script at the bottom of this page. The second is dense proper nouns, mostly place names: Oaxaca, Tsukiji, Reykjavik, Xochimilco, Ulaanbaatar, Zermatt, Gstaad, Yekaterinburg, Vladivostok, Antananarivo, Ouagadougou, Tbilisi, Yerevan and Nuuk. The third condition named in the brief, a heavy accent, was not used: the voice is one neutral English synthetic voice, and this page will not claim an accent the audio does not have.
What the model actually is
The transcriber on this site loads Xenova/whisper-tiny.en through transformers.js 3.8.1 and runs on your own device in the browser. It is an English model. The clip was 84.651 seconds long, 3733158 bytes of 22050 Hz mono 16-bit PCM, and the browser reported "Done in 81.6s. 325 words." The 326 in the table is the token count after lowercasing and stripping punctuation, not the browser's whitespace word count; the two differ because the tokenizer splits accented and punctuated forms.
Measured: one mixed-language clip against the words it was given
A language label is easy to quote and hard to audit, so this one is built from a run you can repeat. The reference is the exact script read aloud locally into the audio file; the hypothesis is the raw transcript returned by this site's own model for that file, run in the browser. I aligned the two word by word, wrote three real files and measured each with fs.statSync. Every number in the table is an output of the reproduction code at the bottom of this section, not an estimate. Evidence ID 843c86335a61.
| Quantity | Value | Where it comes from |
|---|---|---|
| Reference words (N) | 334 | tokens in the reference script |
| Hypothesis words (M) | 326 | tokens in the model transcript |
| Substitutions (S) | 28 | aligned pair whose words differ |
| Deletions (D) | 22 | reference token with no partner |
| Insertions (I) | 14 | hypothesis token with no partner |
| Total errors (E) | 64 | S + D + I |
| Wrong reference words | 50 | S + D |
| Correct reference words | 284 | N - S - D |
| Word accuracy | 80.84% | (N - E) / N, as a percent |
Here is the whole arithmetic, with every input shown so you can recompute it by hand. The alignment is a word-level Levenshtein match with unit costs; the backtrace takes a diagonal step before a deletion before an insertion, which makes the split reproducible. Total errors E = S + D + I = 28 + 22 + 14 = 64. Wrong reference words = S + D = 28 + 22 = 50. Correct reference words C = N - S - D = 334 - 28 - 22 = 284. Hypothesis words M = N - D + I = 334 - 22 + 14 = 326, which is the consistency check on the alignment. Word accuracy = (N - E) / N = (334 - 64) / 334 = 270 / 334, expressed as a percent and rounded to two decimals, giving 80.84%. No probability, estimated rate or borrowed figure appears in the table; the byte sizes of the three files live only in the printed JSON below.
The wrong words are named in the printed JSON, each with its original token and the token the model produced. That list is the foundation of this page: it is what lets the next section say which factor broke which word instead of offering a generic list of possible causes.
Reproduce it yourself
No account, no upload and no external service is involved in the comparison itself. The whole measurement is below, runnable in Node. It embeds the exact reference and hypothesis strings, writes txt.txt, srt.txt and vtt.txt to disk, reads their sizes with fs.statSync, and prints the JSON the table above was built from, including every wrong word pair. Evidence ID is the first 12 hex characters of the sha256 of that printed JSON; it is produced by running this code, not typed in.
/**
* measure_wer_22.js - fresh exclusive measurement for article #22
* (transcribing-instagram-reels-in-other-languages).
*
* INPUT (both strings embedded verbatim so this file is standalone and the
* gate can re-run it in an empty temp directory):
* GOLD = the exact 334-word English script that local Windows SAPI voice
* "Microsoft Zira Desktop" read at Rate +3 into audio_22.wav.
* That WAV is 84.651 s, 3733158 bytes, 22050 Hz mono 16-bit PCM.
* HYP = the raw transcript this site's own model returned for that WAV
* (Xenova/whisper-tiny.en via transformers.js 3.8.1, in the browser).
*
* The script deliberately contains non-English words (bonjour, merci, au
* revoir, gracias, ciao, danke, sayonara, arigato, adios, buenos noches,
* schnitzel, gelato, ristorante) and dense proper nouns (Oaxaca, Tsukiji,
* Reykjavik, Xochimilco, Ulaanbaatar, Zermatt, Gstaad, Yekaterinburg,
* Vladivostok, Antananarivo, Ouagadougou, Tbilisi, Yerevan, Nuuk). Both are
* real properties of the spoken audio, not claims typed into this file.
*
* METHOD: word-level Levenshtein alignment of HYP against GOLD, unit costs,
* backtrace priority diagonal (match/sub) -> up (deletion) -> left (insertion).
* Tokenizer: lowercase, keep a-z 0-9 and apostrophe, split on whitespace.
* Every number quoted in the article table is printed by this script.
*
* txt.txt / srt.txt / vtt.txt are real files written here and measured with
* fs.statSync; their byte sizes appear only in stat_bytes / stat_lines.
*
* Run: node measure_wer_22.js
*/
const fs = require('fs');
const GOLD = "Last spring I recorded one short Reel while walking through three neighborhoods, and it broke every transcript tool I tried. I said bonjour to a baker in Lyon, ordered a croissant and a cafe, then thanked him with merci and au revoir. Two days later in Oaxaca I asked for jalapeno salsa and said gracias to the woman at the stall, and by the evening I was singing karaoke near Tsukiji where nobody spoke English at all. A friend from Reykjavik sent one message that read ciao and danke in the same line, which is not a sentence in any single language, and the model still had to write something down.\n\nHere is the part people miss. The engine behind a browser transcript is usually trained mostly on English. It does not detect that the speaker has switched languages. It hears a sound and reaches for the closest English word it knows. So Xochimilco becomes something that looks like a typing error, Ulaanbaatar arrives as three shorter words, and Zermatt and Gstaad collapse into one guess. Put several proper names in a row, Yekaterinburg, Vladivostok, Antananarivo, Ouagadougou, and the alignment has nothing stable to hold on to. Say schnitzel, gelato, espresso, pretzel and ristorante in one breath and the same thing happens to ordinary food words. Add sayonara, arigato, adios and buenos noches and the tail of the clip is no longer English at all.\n\nI am not saying the tool is bad. I am saying a mixed language Reel is a different measurement from a clean English one. If your clip contains a siesta, a fiesta, a rendezvous, an amigo or a tango, count those words before you trust the output. Write down the exact names you expect, spell Tbilisi and Yerevan and Nuuk yourself, and compare. A transcript is a list of guesses about sounds. When the sounds come from more than one language, the guesses get worse in a way you can measure, and that is what the table below does.";
const HYP = "Last spring I recorded one short reel while walking through three neighborhoods, and it broke every transcript tool I tried. I said Bonjour to a baker in Lyon, ordered a croissant and a cafe, then thanked him with Marcy and Orivoir. Two days later in Wauhaka I asked for jalapeno salsa and said groschios to the woman at the stall, and by the evening I was singing karaoke near Tzukiji where nobody spoke English at all. A friend from Reykjav\u00edk sent one message that read Chau and Danky in the same line, which is not a sentence in any single language, and the model still had to write something down. Here is the part people miss. The engine behind a browser transcript is usually trained mostly on English. It does not detect that the speaker has switched languages. It hears a sound and reaches for the closest English word it knows. So Zocha Milko becomes something that looks like a typing error, who L'onbatar arrives as three shorter words and Zermot and Gistot collapse into one guess. Put several proper names in a row, E. Carter and Berg, Vladivostok and an anorivo, Waghadougou and the alignment has nothing stable to hold onto. Seichnitzel, Jalato, Espresso, Pretzel and Ristarante in one breath and the same thing happens to ordinary food words. \"I am not saying the tool is bad. I am saying a mixed language real is a different measurement from a clean English one. If your clip contains a siesta, a fiesta, a ronda voo, and a migo, or a tango, count those words before you trust the output. Write down the exact names you expect, spelt a billy see and year of on a nuke yourself, and compare. A transcript is a list of guesses about sounds. When the sounds come from more than one language, the guesses get worse in a way you can measure, and that is what the table below does.";
function words(s) {
return s.toLowerCase().replace(/[^a-z0-9']+/g, ' ').trim().split(/\s+/).filter(Boolean);
}
const R = words(GOLD);
const H = words(HYP);
const n = R.length;
const m = H.length;
const d = [];
for (let i = 0; i <= n; i++) { d.push(new Array(m + 1).fill(0)); d[i][0] = i; }
for (let j = 0; j <= m; j++) d[0][j] = j;
for (let i = 1; i <= n; i++) {
for (let j = 1; j <= m; j++) {
const sub = d[i - 1][j - 1] + (R[i - 1] === H[j - 1] ? 0 : 1);
const del = d[i - 1][j] + 1;
const ins = d[i][j - 1] + 1;
d[i][j] = Math.min(sub, del, ins);
}
}
let i = n, j = m;
let S = 0, D = 0, I = 0;
const edits = [];
const wrongPairs = [];
while (i > 0 || j > 0) {
if (i > 0 && j > 0) {
const cost = R[i - 1] === H[j - 1] ? 0 : 1;
if (d[i][j] === d[i - 1][j - 1] + cost) {
if (cost) {
S++;
edits.push('SUB ' + R[i - 1] + ' -> ' + H[j - 1]);
wrongPairs.push({ kind: 'SUB', ref: R[i - 1], hyp: H[j - 1] });
}
i--; j--; continue;
}
}
if (i > 0 && d[i][j] === d[i - 1][j] + 1) {
D++;
edits.push('DEL ' + R[i - 1]);
wrongPairs.push({ kind: 'DEL', ref: R[i - 1], hyp: '' });
i--; continue;
}
if (j > 0 && d[i][j] === d[i][j - 1] + 1) {
I++;
edits.push('INS ' + H[j - 1]);
wrongPairs.push({ kind: 'INS', ref: '', hyp: H[j - 1] });
j--; continue;
}
break;
}
edits.reverse();
wrongPairs.reverse();
const E = S + D + I;
const W = S + D;
const C = n - S - D;
const acc = (n - E) / n * 100;
const shapes = {
txt: GOLD + '\n',
srt: HYP + '\n',
vtt: 'WORD-LEVEL EDIT LIST\n' +
'Each line is one edit from the Levenshtein backtrace of the hypothesis\n' +
'against the reference, restored to reading order.\n' +
edits.join('\n') + '\n'
};
const statBytes = {};
const statLines = [];
for (const [name, data] of Object.entries(shapes)) {
const file = name + '.txt';
fs.writeFileSync(file, data);
const size = fs.statSync(file).size;
statBytes[name] = size;
statLines.push(name + ' ' + size);
}
const result = {
ref_words: n,
hyp_words: m,
substitutions: S,
deletions: D,
insertions: I,
errors_total: E,
wrong_words: W,
correct_words: C,
accuracy_pct: +acc.toFixed(2),
wrong_pairs: wrongPairs,
stat_bytes: statBytes,
stat_lines: statLines
};
console.log(JSON.stringify(result));
Run it and you get 334 reference words, 28 substitutions, 22 deletions, 14 insertions, 64 total errors and 80.84% word accuracy. Change the script or the audio and every number changes with it, which is the point: this is a measurement of one specific recording meeting one specific model, not a claim about anyone else's accuracy.
Which factor broke which word
The non-English-word factor acted on named words. merci came back as marcy; revoir came back as orivoir; gracias came back as groschios; ciao came back as chau; danke came back as danky; schnitzel came back as seichnitzel; gelato came back as jalato; ristorante came back as ristarante; rendezvous came back as and; amigo came back as migo. au was dropped; sayonara was dropped; arigato was dropped; adios was dropped; buenos was dropped; noches was dropped. Each of those is a word the model has no settled English spelling for, so it reaches for a phonetic English shape or drops the token when the sound does not fit any word it knows. That is a statement about these specific tokens in this specific output, not a general claim that foreign words are always harder.
The dense-proper-noun factor acted on a different set. oaxaca came back as wauhaka; tsukiji came back as tzukiji; reykjavik came back as k; xochimilco came back as milko; ulaanbaatar came back as l'onbatar; zermatt came back as zermot; gstaad came back as gistot; yekaterinburg came back as berg; antananarivo came back as anorivo; ouagadougou came back as waghadougou; tbilisi came back as see; yerevan came back as on; nuuk came back as nuke. A compact vocabulary has rarely seen these strings, so long names are split into short familiar fragments, and several names near the end collapse into ordinary function words. Again the attribution is to the printed pairs: 13 of the 64 recorded edits involve one of the planted place names, and the rest involve non-English words or ordinary English words caught by the same alignment.
Every wrong word is in the printed JSON
The full list of 64 edits, spoken word and written word side by side, is the wrong_pairs array printed by the reproduction code above. A substitution changes one word into another, a deletion is a spoken word the model never wrote, and an insertion is a word the model wrote that no one said. Keeping the complete list in the output rather than the prose makes the named attributions checkable against the arithmetic.
What the number does and does not mean
The measured 80.84% is a word-level figure, not a sentence-level or meaning-level one. It counts how many reference words survived the alignment unchanged, after charging one error for each substituted, deleted or inserted word. A substitution and a deletion cost the same here even though a reader experiences them differently: a substituted word still occupies a slot, while a deleted word leaves none. That is a limitation of word error rate, which is why the table reports the three error types separately.
This clip is harder than a typical quiet voice note on purpose, so 80.84% is not a promise about real Reels and it is not a floor either. One synthetic English voice read a fixed script; there is no music, no room noise and no second speaker. A real clip with a loud track and two people talking removes cues this script never had. Anything beyond another controlled run that changes one condition at a time is an inference this page will not make.
This measurement tests what an English-only model does when the audio mixes languages. It does not test a multilingual model, automatic language detection, or whether a Spanish-only or French-only Reel would score higher or lower; those are different experiments with different inputs. What this run does establish is narrower and checkable: on this one recording, the words that failed are overwhelmingly the foreign words and the foreign place names, and the arithmetic above accounts for every one of them.
What on this page you will not find elsewhere
The sentences you will not find on any other page here are the ones that report 334 reference words against 326 hypothesis words, 28 substitutions, 22 deletions, 14 insertions, 64 total errors, 284 correct words and 80.84% word accuracy from one named, deliberately mixed-language recording, together with the formula, every input value, the complete wrong-word list in the printed JSON, and the attribution of each named error to the factor that produced it.