Best real-time speech translation APIs in 2026

Real-time speech translation APIs take live audio in one language and return a translation while the conversation is still going. Some return translated text you show as captions. Some return translated speech. A few return both. They also differ on how many language pairs they cover, how long you wait per turn, how they bill, and whether the product is still in preview.

This guide compares ten APIs on those criteria, using figures from each vendor's own docs and pricing pages as of November 2026. If you only need transcription, not translation, see our comparison of the best APIs for real-time speech recognition and transcription.

How we compared real-time speech translation APIs

Six criteria decide whether a translation API fits a product. Each one below comes with the question to ask the vendor.

• Language pairs. How many languages does the API translate between, and is every source paired with every target or only to and from English? A vendor's language count often covers transcription, not translation. Look for the translation table.
• Per-turn latency. How long after a speaker finishes a phrase does the translation arrive? Few vendors publish a millisecond figure. Where they don't, we say so instead of estimating.
• Speech-to-speech vs captions. Does the API return translated text, translated audio, or both? Captions on a screen and a dubbed voice in an earpiece need different outputs.
• Pricing model and unit. Is it billed per minute of audio, per hour of session, per token, or as an add-on to transcription? Units differ across vendors, so convert to your own traffic before comparing.
• GA vs preview. Is the product generally available, or labeled preview? Preview models can change their interface or pricing.
• Deployment and regions. Where does audio get processed, and is there an EU option for the translation path specifically?

Real-time speech translation APIs at a glance

As of November 2026. "Not published" means we couldn't find the figure on the vendor's official pages.

The sections below cover each API in turn, with limits stated alongside features.

1. AssemblyAI: Universal-3.6 Pro Realtime with LLM Gateway translation

AssemblyAI does real-time translation inside the streaming speech-to-text connection. You set the llm_gateway connection parameter when you open the WebSocket. When a turn ends, the server sends the final transcript to LLM Gateway with your translation prompt. The translation comes back on the same WebSocket as an LLMGatewayResponse message. There's no second API to call.

Transcription runs on Universal-3.6 Pro Realtime (universal-3-6-pro). It's a streaming-only model covering 32 languages with native code-switching, so a speaker can switch languages mid-sentence and the transcript follows. The real-time translation docs show the setup as one extra connection parameter:

import json

LLM_GATEWAY_CONFIG = {
    "model": "qwen3.5-4b-32k-fast",
    "prompt": "Translate the following English into natural Spanish and reply with only the translation: {{turn}}",
}

CONNECTION_PARAMS = {
    "sample_rate": 16000,
    "speech_model": "universal-3-6-pro",
    "llm_gateway": json.dumps(LLM_GATEWAY_CONFIG),
}

• Language pairs: There's no published list of translation pairs. You set the target language in the prompt. In practice the source side is bounded by the 32 languages the model transcribes (listed in the multilingual transcription docs) and the target side by what the chosen LLM can write. For translating pre-recorded files instead, see the translation feature page. The prompt can also include up to the last ten translated turns as context, and the docs call history "the main lever for translation quality".
• Latency: The docs say that with a small model such as qwen3.5-4b-32k-fast, the translation "typically lands within about 100 ms of the final turn." For the transcription step, Universal-3.6 Pro Realtime posts a 307 ms median time to final segment on Pipecat's STT benchmark. AssemblyAI doesn't publish an end-to-end speech-in to translated-speech-out figure.
• Output: Translated text, delivered per final turn. Partial turns stay in the spoken language. Translations of short consecutive turns can arrive out of order, so you match them by turn_order. If a gateway call fails or times out, that turn gets no translation message.
• Speech-to-speech: Not built in. To speak the translation, you add a separate step with AssemblyAI-TTS-1, which covers 6 languages: English, Spanish, French, German, Italian and Portuguese. There's no documented speech-to-speech translation recipe, so that pipeline is yours to build.
• Pricing: Billed additively. The streaming session costs $0.45/hr of WebSocket-open time, and each translation is billed as LLM Gateway input and output tokens. The docs' example model, qwen3.5-4b-32k-fast, is $0.10 per 1M prompt tokens and $0.50 per 1M completion tokens. The default if you omit a model, gemini-2.5-flash-lite, is $0.10 and $0.40. See the pricing page and the LLM Gateway model list.
• Status: Documented in the streaming docs with no beta or preview label. Translation requests go to LLM Gateway with zero data retention.
• Regions: Streaming has US and EU data-residency endpoints. The docs don't say whether Universal-3.6 Pro Realtime or the llm_gateway translation path runs on the EU streaming endpoint.

Best for: Teams that need captions or translated text alongside an accurate source transcript on one connection, and want to choose the translation model and prompt themselves.

2. OpenAI gpt-realtime-translate

OpenAI's gpt-realtime-translate has its own endpoint, /v1/realtime/translations, over WebRTC or WebSocket. It returns translated audio plus transcript deltas, and the guide says output arrives "while the speaker is still talking".

• Language pairs: Not published on the guide or model page.
• Latency: No millisecond figure.
• Output: Speech-to-speech plus text. You need one session per output language, so translating into three languages means three sessions.
• Pricing: $0.034 per minute, priced by audio duration rather than text tokens.
• Status: No GA or preview label shown.
• Constraints: WebSocket audio is base64-encoded 24 kHz PCM16.

Best for: One-to-one speech-to-speech translation, where you can accept an unpublished language list and per-minute billing per output language.

3. Google Gemini Live API translation

Google's Live API translation runs on gemini-3.5-live-translate-preview. It takes a continuous audio-only stream and returns translated speech, with optional input and output transcripts.

• Language pairs: The prose says "70+ languages". The supported-languages table lists 78.
• Latency: Described as "low-latency", with 100 ms input chunks recommended. No millisecond figure.
• Output: Speech-to-speech. Translation mode doesn't support tools or instructions.
• Pricing: On the paid tier, input is $3.50 per 1M tokens or $0.0053/min of audio, and output is $21.00 per 1M tokens or $0.0315/min. Google gives an effective price of about $0.0368 per minute. A free tier is available.
• Status: Preview.
• Known limits (from Google): Voice replication "can be inconsistent". Language detection struggles with heavy accents, similar languages and rapid language switches.

Best for: Speech-to-speech translation prototypes across the 78 languages in Google's table, where a preview model is acceptable.

4. Microsoft Azure Speech translation

Azure Speech translation comes in two forms. Standard real-time translation returns text, synthesized speech or both. Live Interpreter does speech-to-speech with automatic language identification, language switching and personal voice.

• Language pairs: 91 languages for translation to text. The speech-to-speech target count isn't published.
• Latency: Described as "low latency". No millisecond figure.
• Output: Text, speech or both.
• Pricing (Retail Prices API, East US): S1 Speech Translation is $2.50 per hour and covers up to 2 target languages. More targets add Translator fees. Live Interpreter is billed per component: input audio at $1.00/hr, standard audio output at $1.50/hr, custom audio at $2.00/hr, and text output at $10 per 1M characters.
• Status: The overview has no preview label. Live Interpreter's status isn't stated.
• Regions: Priced and deployed per Azure region.

Best for: Teams already on Azure that need text translation across many languages, or speech-to-speech through Live Interpreter, within their existing cloud region.

5. Speechmatics

Speechmatics translation works in batch and real time. It isn't available in Speechmatics' Agent STT. Translations arrive as AddTranslation and AddPartialTranslation messages.

• Language pairs: "Over 30 languages". The list has 34, all to and from English. Up to 5 target languages per session.
• Latency: Partials give lower latency. The docs warn that "Realtime sessions may have a 5-second delay when finalizing translations."
• Output: Text only, in JSON output.
• Pricing: $0.65/hr translation add-on on top of real-time transcription ($0.45/hr Standard, $0.80/hr Enhanced), so about $1.10/hr or $1.45/hr in total.
• Status: No preview label.

Best for: English-centric translated captions where a delay of a few seconds on final translations is acceptable.

6. Soniox

Soniox real-time translation returns transcript and translation in one token stream. The docs say translation can begin before the sentence ends.

• Language pairs: "60+ languages". The table lists 60, and translation "works between any pair of supported languages". There are one-way and two-way modes.
• Latency: No millisecond figure.
• Output: Text tokens. No speech output.
• Pricing: Token-based. In real time, audio input is $2.00 per 1M tokens and text output is $4.00 per 1M tokens, which covers transcription and optional translation.
• Status: No preview label.

Best for: Two-way conversational captions between non-English pairs, delivered as text.

7. Gladia

Gladia translation works on pre-recorded and live audio. It returns a translated transcript, subtitles and sentences.

• Language pairs: "100+ languages", per its pricing page.
• Latency: Not published.
• Output: Text.
• Pricing: 0.05€/hr add-on (0.0008€/min).
• Status: No preview label.

Best for: Adding translated text to an existing Gladia transcription pipeline, where you can test latency yourself.

8. DeepL Voice API

The DeepL Voice API streams over WebSocket. It returns source transcripts, text translations into multiple targets, and translated speech for a subset of languages.

• Language pairs: The Voice API table lists 48 languages for translation, and "all source languages can be translated into any target language". Translated speech covers only some of them. The DeepL Voice product page says "40+ languages".
• Latency: DeepL recommends 50–250 ms audio chunks for low latency. It doesn't publish a result latency.
• Session limits: Up to 5 translation targets, 1 translated-speech target per session, a 30-second no-data timeout, and audio sent at no more than 2x real time.
• Pricing: The Voice API price isn't published. DeepL Pro plans include 30 minutes of DeepL Voice per user per month.
• Status: Available to customers with a paid DeepL API subscription.

Best for: Teams already on DeepL for text translation that want one text-to-many-targets stream with optional translated speech.

9. Palabra

Palabra offers a speech-to-speech translation API. It returns translated speech, with text for captions or transcripts alongside.

• Language pairs: The pricing page says "60+ languages" for speech-to-speech. The homepage says "70+".
• Latency: Palabra claims under one second. No measured breakdown is published.
• Output: Speech-to-speech plus text.
• Pricing: Pay-as-you-go at $0.04 per minute for speech-to-speech (pricing page).
• Status: No GA or preview label.
• Deployment: Palabra says the API runs in multiple regions and can be deployed in a private cloud or on-premises. It doesn't list the regions.

Best for: Speech-to-speech translation that needs a private-cloud or on-premises option.

10. Krisp Voice Translation API

The Krisp Voice Translation API does real-time speech-to-speech translation over WebSocket, with Python and JavaScript SDKs.

• Language pairs: 61 languages, any-to-any.
• Latency: Krisp claims sub-second latency. No figure is published.
• Output: Speech-to-speech.
• Pricing: Krisp's pricing page lists the Voice Translation API at $4.00/hr pay as you go, or $3.00/hr on the $400/month Growth plan. The API product page lists subscription plans for its krisp-vt-pro model starting at $249/month with 45 hours included and $7.00/hr overage. Enterprise is custom.
• Status: Labeled "New" in Krisp's navigation. No GA or preview label.
• Regions: Not listed. Krisp says no voice data is stored on its servers.

Best for: High-volume speech-to-speech translation with predictable monthly hours.

How to choose a real-time speech translation API

Start from the output your users need, then narrow by the other criteria.

1. Decide between captions and speech. If people read the translation on a screen, glasses display or subtitle track, a text-output API works, and you can show the source transcript too. If people hear the translation, look at the speech-to-speech APIs (OpenAI, Gemini, Azure Live Interpreter, Palabra, Krisp, DeepL on a subset), or add a TTS step to a text API. If you build your own pipeline, the target language has to be one your TTS speaks.
2. Check your exact language pairs. Pull the vendor's translation table, not its transcription table. Some APIs only translate to and from English. Some publish no list. If your users switch languages mid-sentence, test code-switched audio specifically.
3. Measure per-turn latency yourself. Only a few vendors publish a millisecond figure, and none publish a full speech-in to speech-out number under your network conditions. Time it from end of speech to translation arriving, on your own audio.
4. Convert pricing to your traffic. Per-minute, per-hour, per-token and subscription pricing don't compare directly. Model a typical session, including how many target languages you need. OpenAI uses one session per output language, and Azure's S1 rate covers two targets.
5. Check release status and regions. Preview models can change. If you need EU processing, confirm that the translation path itself runs in the EU, not just the transcription.
6. Look at the source transcript. Translation quality depends on what the model heard. Names, numbers and product terms that are wrong in the transcript stay wrong in the translation. The AssemblyAI benchmarks page shows how Universal-3.6 Pro Realtime does on entities and code-switching.

For the architecture behind these choices, including cascaded vs end-to-end pipelines, partial vs final turns and a per-stage latency budget, see our guide to real-time speech translation.

Frequently Asked Questions (FAQs)

What is the best real-time speech translation API?

It depends on the output you need. For translated speech, OpenAI gpt-realtime-translate, Gemini Live API translation (preview), Azure Live Interpreter, Palabra and Krisp all return audio. For translated captions alongside an accurate source transcript, AssemblyAI, Soniox, Speechmatics, Gladia and DeepL return text. Choose by language pairs, measured per-turn latency on your audio, and price at your volume.

What are the alternatives to OpenAI gpt-realtime-translate?

For speech-to-speech output, the alternatives are Google's Gemini Live API translation, Azure Speech translation with Live Interpreter, Palabra, Krisp, and DeepL Voice for a subset of languages. If you only need text, AssemblyAI (Universal-3.6 Pro Realtime with LLM Gateway), Soniox, Speechmatics and Gladia stream translated text. OpenAI publishes no language list and needs one session per output language.

How much does a real-time speech translation API cost?

Units vary. As of November 2026, OpenAI charges $0.034/min, Gemini about $0.0368/min effective, Palabra $0.04/min and Azure S1 $2.50/hr. Speechmatics ($0.65/hr) and Gladia (0.05€/hr) are add-ons to transcription. Krisp lists $4.00/hr pay as you go. AssemblyAI bills $0.45/hr of streaming plus LLM Gateway tokens per translated turn. Soniox bills per token, and DeepL doesn't publish Voice API pricing.

Which real-time translation API supports the most languages?

By published translation counts, Gladia lists "100+" languages, Azure lists 91 for translation to text, and Gemini's table lists 78. Krisp lists 61, Soniox 60 (any pair), DeepL Voice 48 and Speechmatics 34 (to and from English). OpenAI publishes no count. AssemblyAI transcribes 32 languages in real time and sets the target language in the prompt, but publishes no pair list.

Should I use a speech-to-speech translation API or speech-to-text plus an LLM?

A speech-to-speech API returns translated audio in one call, but you get less control over the source transcript, the translation model and the voice. A cascaded pipeline (streaming speech-to-text, then LLM translation, then optional TTS) lets you inspect and log the transcript and swap the model or prompt. It also gives you text captions with no extra work. The cost is that you assemble and time the stages yourself.

Does AssemblyAI support real-time translation?

Yes. On Universal-3.6 Pro Realtime, you set the llm_gateway connection parameter, and each final turn is translated through LLM Gateway and returned on the same WebSocket as an LLMGatewayResponse. Partial turns aren't translated. Output is text. To speak the translation, you add AssemblyAI-TTS-1, which covers English, Spanish, French, German, Italian and Portuguese.