A speech engine may recognize dozens of languages across separate recordings and still struggle when one speaker switches languages mid-sentence. For multilingual audio transcription, ask three different questions: which languages can the model recognize, how does it choose a language, and does it preserve switches inside one recording? A single advertised language count cannot answer all three.
This guide compares documented behavior for specific model configurations and gives you a short test that exposes missed switches. For a broader comparison of live and recorded tools, see our multilingual transcription guide.
- Language coverage is a model's supported set. It is not a guarantee that every combination works in one file.
- Automatic language detection can pick the best fit from candidates without transcribing every language spoken.
- Code switching means retaining language changes in the transcript, including changes inside a sentence. Verify the exact model and setting.
- Measure the output at each switch on your own recording; vendor benchmarks cannot replace this check.
Three Specifications to Ask For
Coverage answers whether a model supports a language at all. Detection answers how it chooses among possible languages for a request. Code switching answers whether it can follow changes within that request. These are separate behaviors, and model version matters as much as vendor name.
For example, Google Cloud Speech-to-Text's automatic language recognition guide allows up to three candidate languages and says the service uses the best-fit language from that list. Three here counts candidates for selection; it is not a promise that all three will appear correctly in one transcript. Google's page also limits this feature to specific models and regions.
What the Model Documentation Actually Says
| Model or setting | Documented behavior | Question to test |
|---|---|---|
| Open-source Whisper default transcription | Detects a language from up to the first 30 seconds when none is supplied, then uses that language for decoding. | What happens when the language changes later in the file? |
| Google Cloud Speech-to-Text V2 automatic recognition | Selects a best-fit language from up to three candidates. This is language selection, not a published code-switching guarantee. | Does the output retain both languages at the switch points? |
| AssemblyAI code-switching configuration | Its documented manual language-code setting accepts up to two codes, one of them English; its automatic code-switching setting can detect up to two. | Is your chosen model and language pair covered by this configuration? |
| AssemblyAI Universal-3.5 Pro | Documents native code switching across 18 supported languages. Eighteen describes the supported set, not a stated per-file maximum. | Are all languages in your clip among those 18, and how accurate are the switches? |
| Deepgram Nova-2 multilingual | The documented multilingual mode supports English and Spanish code switching. | Does your recording use this exact pair and mode? |
| Deepgram Nova-3 multilingual | Deepgram documents code switching across a set of ten languages with language=multi. Ten is the model's language set, not a published maximum of languages per recording. | Does the configured mode preserve short switches in your target languages? |
These rows describe different kinds of limits. Do not sort their numbers as though they measured the same thing. In particular, Google's candidate count and the supported sets for Universal-3.5 Pro and Nova-3 are not comparable per-file ceilings. OpenAI's hosted file transcription API is also a separate product from the open-source Whisper code linked above; test its current model independently.
How to Test a Mixed-Language Recording
- Make a reference clip. Record two minutes with real speakers. Include each target language, several switches, a mid-sentence switch, and names that matter to your team.
- Keep a manual reference transcript. Mark each language change and proper noun so you can see exactly what was missed.
- Use the exact production settings. Pin the model version and enable the vendor's documented multilingual or code-switching mode. Save the request settings with the result.
- Read every switch point. Look for dropped words, words rewritten in the wrong script, translation when you asked for transcription, and language labels that lag behind speech.
- Repeat with your hardest pair. Add noise or a third language only if those conditions occur in your real recordings. Compare results against the same reference, and arrange human review for consequential uses.
A vendor's accuracy figure may use different languages, speakers, and scoring rules. This procedure tells you whether its output works for your audio. If your workflow also needs translation, our translation tools guide covers that separate step.
Where MirrorCaption Fits
MirrorCaption's live path is built for captions and translation during a conversation. The current client sends language hints with language identification enabled; one-way translation can accept detected source languages supported by the configured service. Automatic detection does not add support for an unsupported language. That configuration is not a measured guarantee for every language pair or every mid-sentence switch. Test a live session with your real speakers before depending on it. Our multilingual call-center guide covers the live workflow.
The current file-upload UI sends one selected source-language hint. Do not infer from the live configuration that uploaded recordings receive the same multilingual handling. For a finished recording with several languages, test the upload result and use a specialist code-switching service or human review if it misses important passages.
Every new account includes one free hour, once, without a credit card. The current Annual plan is €54.99 for 110 hours per year; check current pricing before purchasing.
Frequently Asked Questions
Does a list of 99 languages mean one recording can contain all 99?
No. A supported-language list tells you which languages a model can recognize. Check separate documentation for automatic detection and code switching within one recording.
Is selecting three candidate languages the same as code switching?
No. Google Cloud Speech-to-Text documentation says its automatic recognition selects the best-fit language from up to three candidates. That setting does not promise a mixed-language transcript.
Can open-source Whisper follow a language switch automatically?
The default Whisper transcription path detects a language using up to the first 30 seconds and uses that choice for decoding. Do not assume automatic mid-file switching; test a mixed-language clip or segment it deliberately.
Can MirrorCaption handle a multilingual recording?
MirrorCaption live sessions use language identification and language hints. The current file-upload UI sends one selected source-language hint. Test a representative recording before relying on mixed-language upload output.
Make the Model Prove the Claim
Choose a model for the language pair and workflow you actually have. Check its code-switching documentation, save the exact settings, and test a clip containing real switch points. A long language list is a starting filter; the transcript itself is the buying evidence.
Try a live multilingual session
Use your own speakers and language pair. Your first hour is free.
Get Started Free