{"id":2382,"date":"2026-06-06T09:00:00","date_gmt":"2026-06-06T09:00:00","guid":{"rendered":"https:\/\/wiro.ai\/blog\/?p=2382"},"modified":"2026-09-27T18:04:36","modified_gmt":"2026-09-27T18:04:36","slug":"cohere-transcribe-5-speech-to-text-tests-en-es-fr","status":"publish","type":"post","link":"https:\/\/wiro.ai\/blog\/cohere-transcribe-5-speech-to-text-tests-en-es-fr\/","title":{"rendered":"Cohere Transcribe: 5 Speech-to-Text Tests (EN, ES, FR)"},"content":{"rendered":"

Cohere Transcribe<\/strong> was tested with five short speech-to-text checks in English, Spanish, and French. The point was not to measure a leaderboard score from a handful of clips. It was to inspect the details people notice in a transcript: punctuation, numbers, names, accents, and what happens when the selected language does not match the audio.<\/p>\n

The model on Wiro<\/a> is a 2B Conformer encoder-decoder ASR model. Its documentation lists 14 supported languages and says that the caller must choose the language; it does not auto-detect it. That matters when an app receives multilingual recordings. The five checks below preserve the original audio and transcripts, then add the context needed to read the results honestly.<\/p>\n

\n \"Cohere
Generated cover artwork retained from the original post.<\/figcaption><\/figure>\n

What this Cohere Transcribe test checked<\/h2>\n

Three clips were short text-to-speech recordings. They supplied a known reference sentence in English, Spanish, or French. One English recording was submitted twice: once with language=en<\/code> and once with language=es<\/code>. A fifth audio sample had no supplied reference, so it is included as a transcription example rather than an accuracy claim.<\/p>\n

Every run used maxNewTokens=256<\/code>. That is the default documented on Wiro and leaves enough room for these short clips. The archived outputs do not retain task receipts, so there is no trustworthy per-output runtime or cost to report for Tests 1-5. Wiro task details can expose elapsed time and cost for a run, but those values were not stored with these published examples. Assigning a number now would be guesswork.<\/p>\n\n\n\n\n\n\n
Setting<\/th>\nValue<\/th>\nWhy it matters<\/th>\n<\/tr>\n
Model<\/td>\ncoherelabs\/cohere-transcribe-03-2026<\/td>\nDedicated audio-in, text-out transcription model.<\/td>\n<\/tr>\n
Language<\/td>\nen, es, or fr<\/td>\nRequired selector; the model documentation says it does not auto-detect.<\/td>\n<\/tr>\n
maxNewTokens<\/td>\n256<\/td>\nDefault limit used for each short-clip output.<\/td>\n<\/tr>\n
Audio formats<\/td>\nMP3 in these tests<\/td>\nWiro also documents direct WAV and MP3 input, with other common formats converted to MP3.<\/td>\n<\/tr>\n<\/table>\n

Five Cohere Transcribe speech-to-text tests<\/h2>\n

1. English short clip with language=en<\/h3>\n

Reference:<\/strong> This is a short test sentence for speech to text. It includes numbers like 42 and 3.14, and a name: Jordan.<\/p>\n

Transcript:<\/strong><\/p>\n

This is a short test sentence for speech to text. It includes numbers like 42 and 314 and a name, Jordan.<\/pre>\n
This is a clean sentence-level result. The words, the number 42, and the name Jordan survive. The useful failure is small but real: 3.14<\/code> becomes 314<\/code>. That changes the value. Anyone transcribing measurements, prices, version numbers, or legal figures should keep numeric QA in the workflow. The punctuation is sensible, but the decimal point is more important than the comma before the name.<\/p>\n
2. The same English clip with language=es<\/h3>\n
Transcript:<\/strong><\/p>\n
This is a short test sentence for speech to text. It includes numbers like 42 and 314 and a name, Jordan.<\/pre>\n
The output matches Test 1 on this tiny, clear recording. That does not prove language detection. It only shows that this English sentence still decoded intelligibly when Spanish was selected. The documented contract remains important: choose the real recording language. A wrong value can matter far more with longer audio, language-specific punctuation, names, or code-switching.<\/p>\n
3. Spanish short clip with language=es<\/h3>\n
Reference:<\/strong> Hola, esta es una prueba corta de transcripcion. Incluye el numero cuarenta y dos y la fecha diecisiete de abril.<\/p>\n
Transcript:<\/strong><\/p>\n
Hola, esta es una prueba corta de transcripci\u00f3n. Incluye el n\u00famero 42 y la fecha 17 de abril.<\/pre>\n
The Spanish output restores accents in transcripci\u00f3n<\/em> and n\u00famero<\/em>, then normalizes spoken quantities and dates into digits. That is usually useful for searchable notes. It is not a verbatim transcription: cuarenta y dos<\/em> becomes 42<\/code>, and diecisiete de abril<\/em> becomes 17 de abril<\/code>. Pick this style for readable meeting notes or search indexes. Pick a review step if the source wording itself must be preserved.<\/p>\n
4. French short clip with language=fr<\/h3>\n
Reference:<\/strong> Bonjour, ceci est un court test de transcription. Il contient le nombre quarante-deux et la date dix-sept avril.<\/p>\n
Transcript:<\/strong><\/p>\n
Bonjour. Ceci est un court test de transcription. Il contient le nombre 42 et la date 17 avril.<\/pre>\n
French follows the same pattern. The model separates the greeting into its own sentence and converts the spoken number and date to numerals. The wording otherwise remains close to the reference. This is a good example of why a transcript should be judged on the intended output format, not only character-for-character identity. For a customer-call summary, normalized figures may help. For a quotation or language dataset, compare against the audio before publishing it.<\/p>\n
5. English sample clip<\/h3>\n
Transcript:<\/strong><\/p>\n
Finally, there are many small cats including loose pet cats that eat the far more numerous small prey like insects, rodents, lizards and birds.<\/pre>\n
No source script was retained for this sample, so it cannot support a word-error claim. It does show a complete grammatical sentence with a list of animals and prey types. Treat it as a usability sample only. That distinction keeps a demo from pretending to be a benchmark.<\/p>\n
How to interpret the outputs<\/h2>\n
These examples point to a practical strength: the outputs are readable without a cleanup pass. Capitals, sentence breaks, accents, and list punctuation appear where a reader expects them. The model also normalizes spoken numerals. That helps when transcripts feed search, CRM notes, or support summaries.<\/p>\n
The same behavior creates the main caution. A normalized value can be wrong in a consequential way. Test 1 lost the decimal in 3.14. Tests 3 and 4 converted words to digits. Neither is automatically bad, but neither is a faithful record of every character spoken. Audio that contains IDs, currency, decimal measurements, email addresses, product codes, or proper names needs a human check or targeted validation.<\/p>\n
The model card says preprocessing resamples audio to 16 kHz where needed and averages stereo inputs to one channel. It also describes chunking and reassembly for long recordings. Those features make the model a better fit for recordings beyond these short clips, but this post does not claim to have tested them. For another perspective on longer-form transcription, see Cohere Transcribe: Speech-to-Text in 7 Audio Tests<\/a>. For a side-by-side speech workflow angle, see Speech-to-Text APIs in 2026<\/a>.<\/p>\n
When to choose Cohere Transcribe<\/h2>\n
Choose Cohere Transcribe on Wiro<\/a> when the recording language is known, the workflow needs one of its 14 supported languages, and readable punctuation matters. It suits meeting recordings, call archives, interviews, and search-oriented audio indexes where normalized dates and numbers are useful.<\/p>\n
Use a verification layer when exact numeric notation or verbatim wording matters. The short English clip makes that case clearly. Split or label recordings before transcription if speakers switch languages, because the language setting is explicit rather than automatic. Keep the selected language beside the saved transcript so later reviewers can reproduce the run.<\/p>\n
For model background, Cohere’s official Transcribe announcement<\/a> describes the open-weight 2B Conformer model, supported languages, and its reported benchmark results. The Hugging Face model card<\/a> documents preprocessing, long-form chunking, punctuation control, and local usage. Run the model on Wiro<\/a> with the correct language selected, then review the fields that cannot afford a transcription mistake.<\/p>\n","protected":false},"excerpt":{"rendered":"
Cohere Transcribe was tested with five short speech-to-text checks in English, Spanish, and French. The point was not to measure a leaderboard…<\/p>\n","protected":false},"author":4,"featured_media":2381,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[52],"tags":[195,196,63,62],"class_list":["post-2382","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-model-reviews","tag-cohere","tag-cohere-transcribe","tag-speech-to-text","tag-text-to-speech"],"_links":{"self":[{"href":"https:\/\/wiro.ai\/blog\/wp-json\/wp\/v2\/posts\/2382","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/wiro.ai\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/wiro.ai\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/wiro.ai\/blog\/wp-json\/wp\/v2\/users\/4"}],"replies":[{"embeddable":true,"href":"https:\/\/wiro.ai\/blog\/wp-json\/wp\/v2\/comments?post=2382"}],"version-history":[{"count":6,"href":"https:\/\/wiro.ai\/blog\/wp-json\/wp\/v2\/posts\/2382\/revisions"}],"predecessor-version":[{"id":4197,"href":"https:\/\/wiro.ai\/blog\/wp-json\/wp\/v2\/posts\/2382\/revisions\/4197"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/wiro.ai\/blog\/wp-json\/wp\/v2\/media\/2381"}],"wp:attachment":[{"href":"https:\/\/wiro.ai\/blog\/wp-json\/wp\/v2\/media?parent=2382"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/wiro.ai\/blog\/wp-json\/wp\/v2\/categories?post=2382"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/wiro.ai\/blog\/wp-json\/wp\/v2\/tags?post=2382"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}