How does Level AI translate non-English conversations for analysis?

This article is for → Admins, Super Admins, Implementers

Overview

Level AI runs most AI analyses on English transcripts. If your customers and agents speak other languages, Level AI can translate transcripts into English during ingestion so sentiment, intent, and other AI services can be applied consistently across regions and queues. You will be able to carry out your monitoring, analysis, and business process improvements even when customers and agents speak in Spanish, German, French, or other languages.

Translations happen as part of conversation ingestion (not live during the call). You provide conversations (voice, chat, email), Level AI converts them to text, translates the text to English, runs redactions (to mask sensitive data), and then runs AI analysis on the English output.

This workflow applies to voice calls (after transcription), as well as chat and email text. Translation is not real-time—it runs after the conversation is ingested.

How does translation-to-English enable consistent QA and analytics?

The following examples show how translation-to-English enables consistent QA and analytics:

• Example 1: Spanish support calls for sentiment tracking
  A customer support queue in Mexico runs primarily in Spanish. Level AI transcribes the call, translates it into English, then runs sentiment detection so QA managers can compare sentiment trends across English and Spanish queues using one common language output.

• Example 2: French chats for intent categorization
  A France-based chat team receives French chats about refund status and order cancellation. Level AI translates chat text into English and then detects intent, making it easier to report on top drivers across multilingual channels.

How does Level AI translate non-English conversations?

Level AI runs a multi-step ingestion pipeline—transcription/ingestion → translation → redaction → analysis—so AI services can run on English text.

1. Transcription/Ingestion (voice only)

   • Audio is converted to text using speech-to-text models (example: Deepgram for foreign languages).

2. Translation (voice, chat, email)

   • All text is translated into English.

3. Redaction

   • Sensitive (PCI & PII) data in the transcripts is redacted (masked).

4. Analysis (on English text)

   • Level AI runs its core AI services (Sentiment or Intent detection, Auto QA, Redactions, etc.) on the English transcript.

What languages are supported for translation?

Level AI supports a mix of in-house and third-party translation tools.

• In-house translation models: Spanish, German

• Google Translate (out-of-the-box): Danish, Swedish, Norwegian, Dutch, French

• Custom support (by configuration): Most other languages validated by Google Translate can be supported by updating your account configuration.

How do I request support for an additional language?

If your language isn’t available out of the box, you can request it so your account is configured to translate it into English.

1. Identify the language(s) you need e.g., Italian, Portuguese.

2. Confirm the channel(s) involved i.e., voice, chat, email.

3. Share implementation details needed for setup:

   • Your conversation source (telephony/chat/email provider)

   • Any constraints (region-specific variants, mixed-language calls)

4. Connect with your Level AI TAM to submit a configuration request to enable the language.

When will translation and analysis be available?

Translation and ingestion take place simultaneously (not during a live interaction). So both will complete around the same time.

• Expect translations to occur during ingestion of conversations.

• AI analysis (for example, sentiment/intent) runs after the transcript has been translated into English.

• Once the above steps are completed, the translated conversation transcripts along with their insights will become available across all workflows and pages in your Level instance.

How are translation costs estimated?

Translation costs are typically estimated by converting talk-time hours into character counts.

FAQs and Troubleshooting

Issue: My conversation is in a supported language, but I don’t see analysis/scoring.
Solution: Confirm the conversation completed ingestion successfully. Translation and analysis run during ingestion (non-real-time). If the conversation is newly ingested, wait for ingestion to complete. If it persists, connect with your Level AI TAM to request an account configuration check.
 

Issue: We need language X, but it’s not on the supported list.
Solution: Request custom language support. Level AI can support most languages validated by Google Translate by updating your account configuration.
 

Issue: Why isn’t translation happening during the live call?
Solution: Translation is currently non-real-time and runs during conversation ingestion.
 

Issue: Our voice calls aren’t being translated, but chats are.
Solution: Voice translation requires successful speech-to-text transcription first. Validate that transcription is running for the voice source.

Translation is also done at a channel level, so even if ASR (Automated Speech Recognition) is successful, translation may not occur. In this case, please reach out your Level AI TAM /Support for more clarity.
 

Issue: Cost estimates don’t match our expectations.
Solution: Translation costs are estimated by converting talk-time hours into character counts. Ask your Level AI team to review the assumptions used (average characters per hour, included channels, and volumes). Please reach out your Level AI TAM for details.