Define the recording
Write down the spoken language or languages, likely accents, audio source, and whether speakers switch languages. Keep a short sample with the noise and vocabulary of real use.
Language guide
How to assess deepgram languages for your audio
When assessing deepgram languages, start with the language actually spoken in your recording, then check the current documentation for the model and workflow you intend to use. A supported language name alone cannot guarantee an accurate transcript.
Three connected voice workflows
Language choice affects the whole workflow, not just a setting on a request. These related guides cover the input and output decisions around it.
-
1
deepgram transcription
See how spoken audio becomes text and which transcript choices deserve a test.
-
2
deepgram voice
Distinguish understanding a spoken language from producing a spoken response.
-
3
deepgram video transcription
Plan for dialogue, captions, and audio quality when speech comes from video.
Where language support reaches its limits
A language listing is a starting point, not a promise that every speaker or recording will work equally well.
- A listed language is not an accuracy score
-
Background noise, distant microphones, overlapping speakers, and unfamiliar names can change results even when the spoken language is supported.
WorkaroundTest representative recordings and review names, numbers, and other high-stakes words by hand.
- One label may not cover mixed speech
-
A recording can switch languages mid-sentence or contain borrowed terms. Do not assume a single-language setting will handle every switch.
WorkaroundTest a mixed-language sample and check whether the current model documentation describes that use case.
- Regional varieties need their own checks
-
An accent or regional vocabulary may behave differently from the examples used to describe broad language support.
WorkaroundInclude speakers from the regions your audience represents in the evaluation set.
- Speech input and speech output differ
-
Recognition support does not establish that a matching synthesized voice, pronunciation, or output language is available.
WorkaroundVerify speech-to-text and text-to-speech capabilities separately before designing a multilingual interaction.
Check support step by step
Use the same short evaluation for each language your project needs, rather than relying on a single demonstration clip.
Verify the current model
Check Deepgram’s current documentation for the particular model and speech workflow. Confirm whether a language must be specified, can be detected, or is not supported for that setup.
Compare against a reference
Run representative audio, compare the result with a human-reviewed transcript, and note errors in names, numbers, and code-switched passages. Repeat when the model or source audio changes.
Apply the check to your audience
The right test depends on what people will do with the result. Each scenario calls for different source audio and a different review standard.
Interview researcher
Recordings include several speakers, regional accents, and specialist terms.
Check language support with real interview excerpts, then inspect speaker turns and terminology in the transcript.
deepgram transcriptionVoice assistant designer
A caller speaks in one language but expects a reply in another.
Validate recognition and spoken output as separate capabilities; success at one does not prove success at the other.
deepgram voiceVideo editor
An event recording contains audience questions and occasional language switching.
Test the extracted audio and review caption timing and mixed-language passages before publication.
deepgram video transcriptionCompare language-selection approaches
These are evaluation approaches, not a claim that every Deepgram model offers both options. Check the controls available for the model you choose.
Specify a language
Use detection, if available
1
Best starting point
Specify a language
The recording’s spoken language is known in advance.
Use detection, if available
The spoken language is unknown when audio arrives.
2
Input decision
Specify a language
Choose a supported language setting for the selected model.
Use detection, if available
Verify that the selected model and workflow offer detection.
3
Potential advantage
Specify a language
Makes the intended language explicit.
Use detection, if available
Avoids requiring a language label before processing.
4
Main risk
Specify a language
A wrong label can misrepresent the recording.
Use detection, if available
A detected label can be wrong or too broad for the task.
5
Mixed-language audio
Specify a language
One setting may not describe every passage.
Use detection, if available
Detection availability alone does not establish reliable handling of switches.
6
Review step
Specify a language
Compare output with a reference sample in the chosen language.
Use detection, if available
Check both the identified language and the resulting transcript.
Continue exploring voice workflows
Take the next step with voice tools
This guide helps you frame a language-support test. The destination is a separate voice product, so check its own capabilities and terms before using it for a project.
- Bring representative audio
- Check the destination’s supported workflows
- Review important output yourself
Language support FAQ
Support depends on the model and speech workflow, and the list can change. Consult Deepgram’s current documentation for the specific model you plan to use rather than relying on a general list or an older example.
Do not assume detection is available in every model or configuration. Check the current documentation for your chosen workflow, then test whether identification is reliable on recordings like yours.
Mixed-language speech needs its own test; support for two languages separately does not establish how well a model handles switching between them. Include real code-switched passages in your sample and inspect the transcript closely.
A broad language label does not guarantee equal accuracy across regional varieties, speakers, or recording conditions. Evaluate audio from the accents and environments that matter to your audience.
No. Recognizing speech and generating speech are different tasks with potentially different models and language coverage. Verify each capability separately if your workflow needs both.