Which APIs combine speech to text and translation?

Some speech APIs do both steps in one service, and others expect you to call a recognizer and then a translator yourself. Azure Speech is the clearest single-service option for live speech. Google Cloud separates the two steps, so a Google pipeline needs two calls. OpenAI and Gemini accept audio and can return translated text, but with narrower scope. The sections below compare them on what they return, how they bill and what they cannot do.

Azure Speech translates live audio in one service

Microsoft's Azure Speech service "supports real-time, multi-language speech to speech and speech to text translation of audio streams" ¹. Interim results arrive while the speaker is still talking. Microsoft says interim transcription and translation results are returned as speech is detected, before the final results ². Build your interface for text that changes, because a translation that looks finished can be corrected a moment later.

Billing counts each target language. Microsoft's overview says you are charged for the speech to text transcription and for the text translation for each target language ³. Two targets cost more than one, so limit the outputs you request.

Google Cloud needs two calls: recognition, then translation

Google Cloud keeps the steps apart. Speech-to-Text returns recognized text, and its overview says each alternative in the response contains a transcript with the recognized text ⁴. The overview does not describe translation, so the Speech-to-Text call alone does not give you a translation.

For the translated text, Google's Cloud Translation Basic API offers "quick, plug-and-play access to the standard Neural Machine Translation (NMT) model" ⁵. A Google pipeline therefore sends the audio to Speech-to-Text, takes the transcript, and sends that text to Cloud Translation. An error in the transcript is passed straight into the translation, as the accuracy question explains.

The two-call design gives you more control. You can correct or review the transcript between the calls, and you can choose the translation model separately. It also means two sets of settings.

OpenAI and Gemini translate audio with a model

OpenAI's speech-to-text guide names whisper-1 as the model to use for translating a completed audio recording into English ⁶. The output is English. Check the tool's current options before you build for another output language.

Gemini's audio documentation lists transcription and translation of speech among the things a Gemini model can do with audio input ⁷. This is a general model, not a dedicated speech API. Google's general audio guide says that for dedicated speech-to-text models with real-time transcription, you should use the Google Cloud Speech-to-Text API ⁸. Gemini also has a live transcription session that streams for up to 10 minutes ⁹, so Google's pages point in different directions. Check both before you choose a product for live captions.

It depends on latency, output language and what you keep

• Live speech needs a streaming service. Azure Speech is built for streams and returns interim results, so it fits live captions and meetings.
• Recordings can use a batch call. A finished file can go to whisper-1 or Gemini in one request, which is simpler when you only need English output.
• Many target languages cost more. Azure bills for each target language ³, so limit the outputs you request.
• Your transcript may need review. The two-call design makes the transcript visible, which helps when a name or a number matters.
• Check the language pair. A service may recognize a language without translating it into your target, so test each direction.

What to do before you choose

1. Decide whether you need live output or a finished file. Live speech points to Azure Speech or Google's streaming recognition. A file can use any of the options above.
2. Decide which output languages you need. If you need English only, whisper-1 is the simplest path. If you need several languages, count the cost of each target.
3. Test with your own audio. Send a few real recordings through each candidate and read the transcript and the translation.
4. Keep the transcript. Store the source text next to the translation so that errors can be traced and corrected.
5. Read the current price pages before you budget. Prices and model names change, so use the vendor's own pricing table.

For the user-facing side of the same two steps, read how to use voice input in Google Translate or the guide to speech to text with translation. The term for the whole process is in the speech translation glossary entry. For how the languages on each service differ, see the guide to dictation languages.