Omanghana
Intron Health has launched Sahara v2.5, an upgraded voice artificial intelligence platform designed to understand how Africans naturally alternate between languages during conversations.
The new release addresses code-switching, a common practice in multilingual societies where speakers combine a local language with English, French or another language within the same sentence or dialogue.
Many global speech-recognition systems struggle with such conversations because they are trained primarily on standardized, single-language datasets. These failures can produce missing words, incorrect transcripts and distorted meanings, particularly in sensitive areas such as healthcare, court proceedings and financial services.
Code-Switching Support Expanded
Sahara v2.5 introduces bilingual speech-recognition models covering 12 African languages, including Hausa, Swahili, Zulu and Luganda.
The models are designed to follow speakers as they move between an African language and English without requiring users to pause, change settings or repeat themselves.
The company said the technology can recognize locally specific accents, pronunciations and expressions that are frequently misunderstood by international voice platforms.
“Africa needs AI built for how Africans really speak,” Intron chief executive Tobi Olatunji said, arguing that users should not have to alter their accents or avoid local expressions to communicate with technology.
Trilingual Model Covers Kinyarwanda, English and French
The platform also features a trilingual automatic speech-recognition model capable of following conversations that move between Kinyarwanda, English and French.
This reflects everyday communication patterns in Rwanda, where all three languages may appear during a single conversation.
Intron describes the system as the first African trilingual speech-recognition model of its kind. The company said it developed proprietary algorithms for the technology and has filed related patent applications in the United States.
Seven Languages Added to Platform
New speech-recognition support has been added for Nupe, Kanuri, Nigerian Fulfulde, Tigrinya, Kikuyu, Dholuo and Somali.
Intron says the additions have increased Sahara’s overall speech-to-text coverage from 56 to 63 languages and dialects, including 31 African languages.
The expansion is intended to make voice-based services accessible to communities that remain poorly represented in mainstream artificial intelligence datasets.
Sahara v2.5 also provides localized text-to-speech and conversational voice-agent capabilities across 13 languages. These systems can generate responses using regional accents, appropriate pacing and local speech patterns.
Offline Models Target Connectivity and Privacy Challenges
In addition to its streaming application programming interfaces, Intron has improved the platform’s speed and latency for live transcription, call centers and automated voice services.
The company also supports offline deployment on dedicated hardware, including Nvidia-based systems.
Offline processing allows hospitals, courts and other institutions to use the technology without sending sensitive information to external cloud servers. It can also support facilities in areas with unreliable or expensive internet connectivity.
At PAMO Clinics in Port Harcourt, Nigeria, an offline medical documentation system reportedly processed more than 2,000 consultations and improved documentation speed by four times.
Intron Reports Lower Error Rate
Intron’s published benchmarks indicate that Sahara v2.5 achieved an average word error rate of 34.3% across 12 code-switched African-language tests.
The company compared that result with a reported 53.8% error rate for Gemini 3.6, representing a relative reduction in errors of approximately 36%.
Word error rate measures how often a speech-recognition system incorrectly transcribes, omits or adds words. A lower percentage indicates stronger transcription performance.
The results should be treated as vendor-reported benchmarks because performance can vary according to the selected audio dataset, accents, recording conditions and versions of competing products tested.
Intron also cited a separate evaluation conducted through Gooey.ai for the Bill & Melinda Gates Foundation and CLEAR Global. It said Sahara performed better than competing systems on five of seven Nigerian languages evaluated, including comparisons involving Gemini and Meta’s Omnilingual model.
Technology Used in Kenyan Healthcare
Intron says Sahara is already being used to reduce the administrative burden on healthcare professionals.
At a medical facility in Nairobi, a 14-minute consultation conducted in a mixture of Swahili and English can reportedly be transcribed and converted into a structured clinical note in less than 30 seconds.
The technology is designed to capture the clinical context of a conversation and organize the information into a medical record, allowing doctors to spend less time writing notes.
Because transcription errors in healthcare can affect diagnoses, prescriptions and treatment decisions, accurate recognition of mixed-language speech is particularly important.
Ogun Courts Expand Transcription System
The Ogun State Judiciary in Nigeria has expanded its use of Intron’s courtroom transcription system from an initial pilot to nine courts.
The system records proceedings in which witnesses and legal practitioners may alternate between Yoruba and English.
According to Intron, proceedings and transcription tasks that previously required more than four hours can now be completed in approximately half that time.
The judiciary plans to extend the technology to all 18 High Courts covered by the project.
Voice Agents Deployed in Financial Services
Digital lender Branch International has also deployed Sahara-powered voice agents for debt collection and customer engagement.
The agents can communicate in local languages, continue conversations after normal working hours and respond to customers who mix languages.
Intron said the system recovered more than ₦1.2 million in delinquent loans within one week, including repayments made during evenings and weekends. The company also claimed that the agents performed strongly on loans overdue by more than 356 days.
These commercial results are based on information provided by Intron and have not been independently audited publicly.
Sahara currently supports production and research deployments for more than 40 enterprise customers across Nigeria, Kenya, South Africa, Uganda, Rwanda and Ghana.
Further details about the platform are available on Intron’s Sahara v2.5 website.
Source: Omanghana
