HUMAIN, a company of Saudi Arabia's Public Investment Fund (PIF), launched HUMAIN Voice on November 19, 2026. The Arabic-first platform puts speech recognition, text to speech and a voice agent behind one API key, and targets Saudi and other Arabic dialects, English, and speakers who switch between Arabic and English mid-sentence. In the same press release, HUMAIN announced HUMAIN-Voice-Lite-80M, an open-weight speech recognition model under Apache 2.0 that was not yet downloadable on launch day.
TL;DR
Products: speech to text (Fast, Batch and Realtime modes), text to speech, and a hosted voice agent that HUMAIN's docs label beta.
Languages: Arabic and English, including code-switching. The press release names the Saudi dialects Najdi and Hijazi plus Gulf and other dialects; HUMAIN's FAQ says more than 15 dialects are in production.
Pricing: a free tier of 5K credits with no card, then plans from 49 to 649 SAR a month (about $13 to $173), per voice.humain.com on November 19, 2026.
Open weights: HUMAIN-Voice-Lite-80M, an 80M-parameter transducer model for CPU and GPU, announced but not on Hugging Face as of November 19, 2026.
Hosting: data stays in Saudi Arabia, according to HUMAIN's FAQ.
What is HUMAIN Voice?
HUMAIN Voice is the speech layer of HUMAIN's AI stack, next to the ALLaM Arabic language model the company already publishes on Hugging Face. HUMAIN frames the launch around a gap: in its words, speech AI has mostly been built for English and other high-resource languages, with Arabic treated as an extension. Its FAQ adds that most models handle Modern Standard Arabic only.
HUMAIN's pitch is that dialect and code-switching are the normal case, not the exception. Dr. Ahmed Ali, VP of HUMAIN Voice, describes them as "fundamental characteristics of speech, not edge cases."
HUMAIN Voice's developer docs name the models behind the platform:
Bayan: the speech recognition family. The default model handles Arabic, English and code-switched audio, and an Arabic-only version is also available.
Nida: an Arabic model for 8 kHz phone recordings, batch only.
Nebula: the single text-to-speech model.
How does HUMAIN Voice work?
HUMAIN Voice offers speech to text in three modes:
Batch processes complete recordings, with options for speaker labels, redaction of personal data, and number and date formatting.
Fast transcribes one short, complete clip, such as a single turn in a conversation.
Realtime streams 16 kHz mono audio and returns partial and final transcripts.

The HUMAIN Voice demo in Batch mode, with the redaction and number formatting options. Screenshot: voice.humain.com, November 19, 2026.
Text to speech runs on Nebula with seven stock voices at launch: four Saudi, one Levantine, one Moroccan and one Egyptian. All seven speak Arabic and English, and the voice switches to Arabic automatically when the text contains Arabic script. The docs list speaking styles (news, storytelling, advertising, call center, conversational) and emotions (happy, sad, angry, fearful, surprised, whisper). Voice cloning from a reference clip and its transcript is open to signed-in accounts.
The HUMAIN Voice agent is hosted by HUMAIN. Developers set the prompt, intro and voice from their backend. Users then join from a browser, or from iOS, Android, Flutter and React Native apps through LiveKit clients. The press release describes it as full-duplex speech to speech that replies in the caller's own dialect. The docs mark it as beta, and the site's own page description lists voice-agent capabilities as "upcoming". For background on the two agent designs, see our explainer on cascaded vs speech-to-speech agents.
SDKs are available for TypeScript and Python, with Go marked "soon". A text-to-speech route that follows OpenAI's speech request format is documented, but the docs say it only runs in the development environment for now.

The developer section on voice.humain.com, with a TypeScript example for live captions. Screenshot: November 19, 2026.
What does the HUMAIN Voice demo show?
The HUMAIN Voice site has a no-signup demo for all three products. Before the first request, it asks users to confirm they are 18 or older and consent to their voice being recorded. Uploaded files must be uncompressed WAV, up to 120 seconds and 5 MB.

The voice agent demo: six preset assistants (general, banking, healthcare, real estate, retail, and food and beverage) and a reply dialect setting (automatic, or always Saudi, Modern Standard Arabic, Gulf, Egyptian, Levantine or Maghrebi). Screenshot: voice.humain.com, November 19, 2026.
What does HUMAIN Voice cost?
HUMAIN Voice uses one credit balance for every API. The free tier gives 5K credits, which HUMAIN equates to about 3 hours of batch transcription or 6.5 minutes of voice agent time. Paid plans are Starter (49 SAR a month), Basic (99), Premium (169), Pro (329) and Scale (649), with custom Enterprise deals that include private deployment.

HUMAIN Voice plans and overage rates. Screenshot: voice.humain.com, November 19, 2026.
Service |
No plan (pay as you go) |
Scale plan |
|---|---|---|
Batch transcription, per audio hour |
1.00 SAR (about $0.27) |
0.75 SAR (about $0.20) |
Realtime transcription, per audio hour |
1.80 SAR (about $0.48) |
1.36 SAR (about $0.36) |
Text to speech, per 1,000 characters |
0.224 SAR (about $0.06) |
0.169 SAR (about $0.05) |
Voice agent, per minute |
0.51 SAR (about $0.14) |
0.38 SAR (about $0.10) |
Source for every row: the voice.humain.com pricing table, November 19, 2026. Each plan keeps its own rate once its monthly credits run out. Dollar figures use the fixed rate of 3.75 SAR per dollar.
What is HUMAIN-Voice-Lite-80M?
HUMAIN-Voice-Lite-80M is an 80-million-parameter, transducer-based speech recognition model for Arabic, English and code-switching, released under Apache 2.0 according to HUMAIN. HUMAIN positions it for setups where cost, latency or on-device deployment matter, on CPU as well as GPU.
The press release says the model will ship with results from the Open Universal Arabic ASR Leaderboard, along with a blog post. On November 19, 2026, HUMAIN's Hugging Face organization listed only ALLaM-7B-Instruct-preview, so there are no public accuracy figures for the model yet.
How does HUMAIN Voice compare with other Arabic speech AI providers?
HUMAIN Voice enters a field of Arabic speech specialists. Some focus on one layer (recognition, voices or agents); others, such as AudarAI and Emotech, cover several.
Company |
Based in |
What it offers |
Source |
|---|---|---|---|
Saudi Arabia |
Speech to text, text to speech, voice agent (beta), open-weight model announced |
HUMAIN press release and docs, November 19, 2026 |
|
UAE |
Arabic speech to text |
Voice AI Space Library |
|
UK |
Arabic and multilingual speech to text, text to speech, voice cloning |
Voice AI Space Library |
|
US |
Emirati Arabic text to speech and custom voices |
Voice AI Space Library |
|
Jordan |
Arabic voice AI and voice agents |
Voice AI Space Library |
|
UAE |
Arabic-first voice agents, on-premises option |
Voice AI Space Library |
|
Saudi Arabia |
Voice and chat agents in Saudi Arabic |
Voice AI Space Library |
|
Saudi Arabia |
Arabic voice agents for inbound and outbound calls |
Voice AI Space Library |
|
Singapore |
Arabic-first speech recognition, text to speech, voice translation and voice agents |
Voice AI Space Library |
Library rows describe each product as the company does; they are not a test of accuracy or price.
What does HUMAIN Voice mean for teams building voice agents?
HUMAIN Voice gives teams serving Arabic-speaking users a new option they can test without a sales call: a free tier, public per-hour rates, and docs with dialect-specific examples. HUMAIN's FAQ pitches in-Kingdom data residency and private deployment to regulated industries; banking and healthcare are two of the demo's preset assistants.
What has HUMAIN not published yet?
Word error rate and latency figures for the hosted Bayan and Nebula models.
The HUMAIN-Voice-Lite-80M weights and their leaderboard results, which would be the first public accuracy figures for HUMAIN Voice's open model. Our explainer on how voice AI benchmarks work covers what to check when they land.
Production status for the voice agent (beta in the docs) and the OpenAI-compatible speech route (development environment only).
Related
Reson8 releases Rumble-2-Flash as open weights under Apache 2.0
IBM open-sources Granite Speech 5.0, a 2B speech-to-text model
Looking for an Arabic speech provider? Compare speech-to-text, text-to-speech and voice agent companies in the Voice AI Space Library.
Sources
HUMAIN, "HUMAIN launches HUMAIN Voice, an Arabic-first speech AI platform built for how people actually speak", press release via Zawya, November 19, 2026: https://www.zawya.com/en/press-release/companies-news/humain-launches-humain-voice-an-arabic-first-speech-ai-platform-built-for-how-people-actually-speak-2405269
HUMAIN Voice, product demo, pricing and FAQ, accessed November 19, 2026: https://voice.humain.com/
HUMAIN Voice docs, "Models, languages, and voices", accessed November 19, 2026: https://docs.voice.humain.com/en/models
HUMAIN Voice docs, "Voice Agent", accessed November 19, 2026: https://docs.voice.humain.com/en/agent
Open Universal Arabic ASR Leaderboard, Hugging Face: https://huggingface.co/spaces/elmresearchcenter/open_universal_arabic_asr_leaderboard
HUMAIN on Hugging Face (humain-ai), checked November 19, 2026: https://huggingface.co/humain-ai



