Azure-Translator-Text-translation

Version: 1

Microsoft•Last updated December 2025

Text translation is a cloud-based, multilingual service that uses neural machine translation models (NMT) and/or large language models (LLM) to translate text from one language to another, supporting 135 languages.

Azure Translator

Azure Translator is part of the Azure AI Foundry tools family and powers many Microsoft products and services used by thousands of businesses worldwide for language translation operations. Azure Translator supports features such as text translation, document translation and custom translator and supports 135 languages . The Text translation API enables robust and scalable translation capabilities with customization features suitable for diverse applications.

Key capabilities

About this model

Azure Translator provides users with flexibility to choose general neural machine translation (NMT), or a list of generative AI large language models (LLMs) at a request level. Azure Translator allows customers to choose a model addressing their diverse needs at request level, providing control to them on quality, cost, latency, feature, and data residency. Using generative AI models, Translator offers new capabilities to produce tone translation, gender translation, and adaptive custom translation.

Key model capabilities

Adaptive custom translation

Fine-tune translation output using few-shot learning.

Secure & Compliant

Enterprise grade security with data privacy & compliance.

Flexible integration

Use REST APIs, SDKs, connectors or containers.

Model selection

Choose between LLMs & NMT models.

Use cases

Outbound translation: A publisher of information provides documents or text in multiple languages, addressing the target audience in the recipient’s language. There are different classes and formats of outbound material, for example, marketing flyers, informational videos, or factory floor manuals. Machine translation is more suitable for some classes than for others. As a general rule, the suitability of machine translation is inversely proportional to the creativity of the content. The translation can be published as web content, an electronic document, or as video subtitles or dubbing or be printed on paper.
Inbound translation: Someone receives information in a foreign language and uses Translator to translate the information into their native language. Examples are websites, product reviews, financial and business reports, or bug reports arriving in a foreign language. The tolerance for translation errors might be higher in this use case, but the translation might induce significant misunderstandings in a non-negligible number of cases. Often in this scenario, a machine translation is better than no translation. An individual or a business could automatically filter or classify to extract information, or to apply other AI techniques on documents from a variety of sources, including foreign language documents. Examples could be media monitoring, multilingual virtual assistants, or e-Discovery. The recipient applies machine translation before passing the document to the automatic analysis. Most of the time, this process is fully automated with no human intervention.
Bidirectional translation: Two or more humans who do not speak the same language employ machine translation in a live chat over instant messaging or in a spoken conversation. For example, a support agent doesn’t speak the same language as the customer seeking help.
Raw translation: Publish the translation as delivered by the machine translation system. This use case carries the lowest cost and comes with a non-negligible error rate. There should be mechanisms in place to react to mistranslations, such as consumer feedback.
Post-edited translation: Publish the post-edited translation, which is the machine translation result corrected by a human reviewer. Human intervention increases the cost over raw translation by a factor of more than 1,000, but it significantly reduces the error rate and improves the fluency and understandability of the translation.

Key use cases

Apps, websites, enterprise systems, and multilingual workflows.

Out of scope use cases

Carefully consider using: Translation of nonprofessionally authored material. Examples include: Colloquial writing, Transcribed speech, Social media chat.
Carefully consider applying human review when sensitive data or scenarios are involved: It's important to include a human in the loop for a manual review when you're dealing with high-stakes scenarios (e.g affecting someone's consequential rights) or sensitive data. Machine translation may make mistakes. Consider carefully when to include a manual review step for certain workflows. For example, translating medical records should include human oversight.
Carefully consider when using for awarding or denying of benefits: Translator was not designed or evaluated for the award or denial of benefits, and use in these scenarios may have unintended consequences. These scenarios include:
Medical insurance: This would include using translated healthcare records and medical prescriptions as the basis for decisions on insurance reward or denial.
Loan approvals: These include translating applications for new loans or refinancing of existing ones.
Legal and regulatory considerations: Organizations need to evaluate potential specific legal and regulatory obligations when using any AI services and solutions, which may not be appropriate for use in every industry or scenario. Additionally, AI services or solutions are not designed for and may not be used in ways prohibited in applicable terms of service and relevant codes of conduct.

Pricing

Pricing for Azure-MT (NMT) model is based on characters. Pricing page
Pricing for LLMs are based on token consumption. Pricing page

Technical specs

Text translation offers three models: Azure-MT (neural machine translation model), GPT 4o and GPT 4o mini. Using LLMs text translation can produce translation specific to a gender or tone. The translation output can be finetuned with your domain data and terminology to produce better quality results based on the use case using adaptive custom translation for LLMs and custom translator for Azure-MT model.

Training cut-off date

For Azure-MT model (default), the model training process is iterative and does not follow a fixed cadence. For LLMs, refer to the respective model deployment details.

Input formats

UTF-8 encoded text.

Supported language

130+ languages are supported including Arabic, Bengali (Bangla), Chinese Simplified, Chinese Traditional, Dutch, English, French, French (Canada), German, Hindi, Italian, Japanese, Korean, Malay, Portuguese (Brazil), Portuguese (Portugal), Punjabi, Russian, Spanish, Tamil, Telugu, Thai, Turkish, Urdu, Vietnamese. Full language support list: Azure Translator language support API reference for supported language list: Reference

Supported Azure regions

See the full list of supported Azure regions for Azure Translator linked here .

Sample JSON response

Sample input

{
  "inputs": [
    {
      "Text": "How are you?",
      "language": "en",
      "targets": [
        {
          "language": "es",
          "deploymentName": "gpt-4o-mini",
          "gender": "female"
        },
        {
          "language": "es"
        }
      ]
    }
  ]
}

Sample output

{
  "value": [
    {
      "translations": [
        {
          "language": "es",
          "instructionTokens": 302,
          "sourceTokens": 4,
          "responseTokens": 6,
          "targetTokens": 3,
          "text": "¿Cómo estás?"
        },
        {
          "language": "es",
          "sourceCharacters": 12,
          "text": "¿Cómo estás?"
        }
      ]
    }
  ]
}

Model architecture

Azure-MT: Transformer-based multilingual architecture optimized for high-quality, context-aware neural machine translation.

Long context

Information not available.

Optimizing model performance

Users can get best performance on translating a text phrase into multiple target languages by making individual requests for each language rather than making a single request for multiple languages. This approach helps users to consume available translations instead of waiting for all translations to return by the system. If your volume of translation is high, switch to higher commitment or volume tiers.