What happened
Portal ERP reports that Nigerian voice-technology company Intron launched Sahara v2.5 on August 28, 2026. The update focuses on recognizing speech in which speakers switch between languages within a sentence or conversation, adds seven named languages, and expands Sahara’s speech and voice-agent coverage across African languages. Portal ERP attributes the product details to Intron and notes that the company is also developing a trilingual Kinyarwanda-English-French model.
Portal ERP reports that Sahara v2.5 is designed for bilingual code-switching, in which a speaker moves between languages during the same sentence or interaction. The release expands Sahara’s overall language coverage to 31 African languages. Its text-to-speech and voice-agent capabilities now cover 13 languages, according to the report. Portal ERP specifically names Nupe, Kanuri, Nigerian Fulfulde, Tigrinya, Kikuyu, Dholuo and Somali among the newly supported languages.
The article says Intron reported an average word error rate of 34.3% across 12 languages, compared with 53.8% for Gemini 3.6. Portal ERP presents that as a 36% relative reduction in errors, but the source does not provide the test prompts, audio composition, sample sizes, confidence intervals, evaluation protocol or independent replication. In separate testing commissioned by the Gates Foundation and CLEAR Global, Intron reportedly ranked ahead on five of seven Nigerian languages evaluated against Gemini and Meta’s Omnilingual model. Those results also remain unverified outside the report.
Portal ERP reports that Intron is developing a trilingual speech-recognition model for Kinyarwanda, English and French. The company says Sahara uses proprietary technology covered by U.S. patent filings, but the source does not identify the filings or explain which technical methods they cover. Intron says its training dataset contains more than 150,000 hours of audio from over 53,000 speakers, spanning 64 languages and more than 500 accents. The report does not describe how speakers were recruited, how consent was obtained, or how the data is governed.
The article describes several commercial deployments. Portal ERP reports that Branch International used Sahara-powered collections agents to recover more than ₦1.2 million in delinquent loans in one week, including payments outside normal business hours, and that Intron said the agents outperformed human agents on loans overdue by more than 356 days. The Ogun State Judiciary reportedly expanded transcription from one pilot court to nine, while Intron said a 14-minute Swahili-English medical consultation in Nairobi could be converted into a structured clinical note in under 30 seconds. Sahara was also reportedly deployed offline on Nvidia hardware at PAMO Clinics in Port Harcourt.
Source details: portalerp.com ↗
Why it matters
Voice systems often perform less reliably when speakers use multiple languages or regional varieties in the same interaction. If Intron’s reported results hold up in independent testing, Sahara v2.5 could improve access to voice-based financial, judicial, medical and business services in settings where cloud connectivity, language coverage or data-governance requirements limit the usefulness of larger general-purpose systems. The report also illustrates how locally collected speech data is being used to build specialized AI products.
The central significance of the release is language fit rather than simply a larger language count. Many speech systems are evaluated on conversations that remain in one language, while everyday speech in multilingual communities may shift between languages, dialects and borrowed terms. A system designed for that behavior could reduce the need for users to change how they speak to accommodate software. The source, however, provides no independent assessment of whether the improvement is consistent across speakers, environments or types of code-switching.
The reported applications involve areas where transcription and voice automation can affect real people. Collections agents may influence borrowers’ interactions with lenders. Court transcription can affect records used in legal proceedings. Clinical-note generation can influence documentation and downstream care if errors are not caught. These uses make accuracy, auditability and human review more important than a single aggregate benchmark. Portal ERP does not report error rates for names, amounts, legal terminology, medical terms or disputed conversations.
Offline deployment is potentially important for organizations facing weak connectivity, data-residency requirements or restrictions on sending sensitive audio to external cloud services. Portal ERP reports that Intron deployed Sahara on Nvidia hardware at PAMO Clinics through a donor-funded project. The source does not specify the hardware configuration, operating costs, latency, security controls, retention period or whether the system can operate fully offline for every supported language. Those details will determine whether the approach is practical beyond individual deployments.
The report also points to the growing importance of African-language data and locally focused model development. Intron says it has more than 40 enterprise customers across six countries and raised $1.6 million in pre-seed funding since 2024. These figures are company claims reported by Portal ERP, not independently audited figures. If the underlying dataset and deployments are representative and responsibly governed, specialized regional systems could broaden the market for voice AI. If not, headline language coverage could obscure uneven performance or unresolved data-rights issues.
What to watch next
The key unanswered questions are whether Sahara v2.5’s reported accuracy advantage generalizes beyond Intron’s tests, how performance varies by language, accent and code-switching pattern, and whether the system is broadly available or limited to enterprise deployments. Future scrutiny should focus on independent evaluations, error patterns, privacy and consent practices for the company’s audio dataset, clinical and legal transcription safeguards, and evidence behind the reported business outcomes.
Independent testing is the clearest next step. Reviewers should seek the full methodology behind the reported comparison with Gemini 3.6 and Meta’s Omnilingual model, including language-by-language scores, audio conditions, code-switching definitions, speaker overlap, transcription conventions and whether systems were tested with comparable prompts and hardware. The five-of-seven result from Gates Foundation and CLEAR Global-commissioned testing should also be examined through a published report or data release. Nothing in the source confirms that such materials are publicly available.
Performance should be assessed by use case, not only by average word error rate. A system may have a favorable overall score while still mishandling names, figures, negation, accents or rare languages. For courts, clinics and financial services, the important measures include the severity of errors, how quickly users can correct them, whether corrections feed back into the system, and whether a human remains accountable for the final record or decision. Portal ERP does not provide those operational measures.
The company’s dataset warrants scrutiny as the product expands. Intron says the collection includes 150,000 hours of audio from more than 53,000 speakers across 64 languages and over 500 accents. Observers should look for information about consent, compensation, anonymization, regional representation, speaker rights, deletion requests and restrictions on secondary use. Patent filings and proprietary technology may also limit outside inspection, making documentation and third-party audits especially important.
The reported deployments provide concrete milestones to verify. Follow-up reporting could establish whether the Ogun State Judiciary reached all 18 high courts, whether the collections-agent results persisted over a longer period, and whether the medical-note workflow receives clinician review before entering patient records. It will also be important to learn whether Sahara v2.5 is available to smaller organizations, at what cost, under what connectivity requirements, and with what support for languages whose performance has not been independently documented.