[Feature]: Add support for new languages such as Hebrew / Arabic

[Amit-Nadam]

Summary

Deepgram currently offers strong multilingual support through models like Nova-2 and Nova-3, but Hebrew and Arabic are not yet supported in the main models (only partially through Whisper batch mode). Adding these languages would significantly expand accessibility for users and developers in the Middle East region who rely on accurate real-time speech recognition.

Problem to solve

Many organizations, research teams, and developers working in the Middle East need high-quality ASR (Automatic Speech Recognition) for Hebrew and Arabic, including regional dialects.
Current alternatives (e.g., Whisper, Google Cloud STT) are not optimized for low-latency or high-accuracy streaming in these languages, limiting their use in Academic and linguistic research. Without native model support, Deepgram users must rely on slower, less accurate batch transcriptions via Whisper — missing out on Deepgram’s strengths in streaming and diarization.

Proposed solution

1. Add Hebrew (language=he) and Arabic (language=ar) support to Deepgram’s main Nova-3 (and future) speech-to-text models.
2. Provide both batch and real-time streaming capabilities, ensuring compatibility with existing Deepgram API parameters.
3. Update the Models & Languages Overview documentation to reflect new language availability and usage examples.
This approach will make Deepgram’s ASR accessible and practical for a broader developer and enterprise audience in the Middle East, especially for real-time transcription and voice analytics.

Alternatives considered

No response

Scope

Python only

Priority

Important

Extra context / links

To support the addition of Hebrew and Arabic to Deepgram’s models, here are some open-source and research projects that already provide high-quality data and models in these languages:

1. Ivrit.ai
2. ArTST
   These open datasets and models show strong community and research interest in Hebrew and Arabic speech recognition. Integrating these languages into Deepgram’s Nova model family would leverage existing resources, accelerate adoption across new regions, and align with Deepgram’s global accessibility goals.

Session ID (optional)

No response

Project ID (optional)

No response

Request ID (optional)

No response

[artur-gas]

Hi @Amit-Nadam - thanks for reaching out!

To reassure you, we’re actively planning Arabic and Hebrew models. Are you looking for the STT model to output in a specific Arabic dialect (e.g., Gulf Arabic?), or do you prefer a standard MSA output?

[Amit-Nadam]

Hi! @artur-gas Thank you for the quick and positive reply!

For Arabic, the primary need is Modern Standard Arabic (MSA), as it is widely used in media, formal speech, and education.

However, support for common dialects would be very valuable, especially Levantine Arabic (spoken in Israel, Lebanon, Syria, Jordan) and Egyptian Arabic, which are common in conversational speech, customer service calls, and online content.

For Hebrew, the priority is Modern Israeli Hebrew, including handling of accents and occasional code-switching with English.

These additions would make Deepgram highly relevant for multilingual applications in the Middle East, including transcription, accessibility tools, and education technology.

Thanks again for considering this feedback!

Do you have an estimated timeline for when these language models might be available?

[artur-gas]

@Amit-Nadam - thank you for the details!

Quick question: are you building an STT → LLM → TTS workflow for Arabic and Hebrew speakers? In other words, is the STT output primarily feeding into a downstream system (like an LLM), or is it meant to be reviewed directly by humans?

On a side note, would love to meet and discuss Arabic and Hebrew STT models if you have some availability coming days

[Amit-Nadam]

@artur-gas I’m planning to use the STT output primarily as input for downstream LLMs and other AI models, rather than for direct human review. This workflow will help with tasks like transcription analysis, summarization, and multilingual AI applications.

Could you share an estimated timeline for when these language models will be available? I am planning to use them in my thesis and this information would greatly help with my research schedule.