Which speech-to-text APIs can identify speakers?

Most of the major hosted speech-to-text APIs can label speakers, and so can some open-source setups built around Whisper. The feature is called speaker diarization. It splits a transcript by who is talking, and it is often sold as a separate option or priced differently from basic transcription. Support also varies by language and by product, so check the exact model you plan to call. The speech-to-text API guide compares the providers on price and accuracy.

Hosted APIs that return speaker labels

Three details in the table matter for a purchase. Google's support depends on language, so test your language before you commit. Amazon bundles the feature into its normal rate, while AssemblyAI bills it on top. OpenAI treats speaker labels as a specialist job, not its default transcription model.

Whisper does not label speakers on its own

Whisper is a speech recognizer, and its output does not say who spoke. To get speaker labels from Whisper, you need a second step. The WhisperX project runs Whisper-based recognition and adds word-level timestamps and speaker diarization to its output ⁸. That is the usual way to add speakers to a Whisper transcript, and it is a self-hosted option, so you pay in hardware and setup time rather than per minute. The guide to running Whisper locally explains the setup, and whether Whisper can identify speakers goes deeper on the question.

Speaker labels are not perfect on every recording

Labels are a prediction, and they make their own errors. Test them on your recordings, not on a demo. Use a recording with two speakers who take turns, and count how often the system switches speakers at the wrong point or assigns a line to the wrong person. Include a stretch where people talk over each other, because that is where labels most often slip. The question on transcribing audio with multiple speakers walks through a practical workflow.

What to do: pick a provider, then test its labels

1. Confirm the language. Check that your language is supported for speaker labels, and do not assume it from the general language list.
2. Decide how you will pay. If you need labels on most files, a base price that includes them, or an add-on you can budget, is the simpler choice. The cheapest speech-to-text API question compares the rates.
3. Test on a real two-speaker recording of at least 10 minutes. Count speaker switches that are wrong, and check whether names or roles map to the right voices.
4. Keep the transcript format in mind. Word-level speaker tags are easiest to edit, and some providers return them in structured JSON.
5. If you must run open models, start with WhisperX, and compare it with a hosted API on the same recording before you decide.