# AI4Bharat and Josh Talks Launch Multimodal Voice of India Evaluation Platform > Source: > Published: 2026-08-04 10:23:51+00:00 # AI4Bharat and Josh Talks Launch Multimodal Voice of India Evaluation Platform AI4Bharat at IIT Madras and Josh Talks launched a broader Voice of India evaluation platform in New Delhi on August 4, extending the project beyond its earlier speech-recognition benchmark. The official site now organizes India-focused evaluation across voice, language, vision, and agent categories, while its mature ASR benchmark covers 15 languages and 536 hours of real-world speech. AI4Bharat at IIT Madras and Josh Talks launched a broader **Voice of India** evaluation platform in New Delhi on August 4, 2026. The Hindu BusinessLine reported that the launch event included S. Krishnan, secretary of India's Ministry of Electronics and Information Technology, and that the platform is intended to evaluate AI systems under Indian linguistic, cultural, and deployment conditions. This launch expands the Voice of India identity beyond the speech-recognition benchmark introduced earlier in 2026. The official platform now catalogs evaluation across voice, language, vision, audio, video, and agent tasks, with some categories live and others explicitly marked as coming soon. ### A broader evaluation layer The official site presents speech-to-text, text-to-speech, speech-to-speech, and text-to-image evaluation alongside planned categories such as translation, legal reasoning, image understanding, tool use, and video tasks. It says generative rankings use blind human judgments, hide model labels until a choice is submitted, and randomize left-right order. Speech recognition uses orthographically informed word error rate, or OI-WER, where lower scores are better. That distinction matters because the platform is not one uniform leaderboard. Human-preference Elo scores and ASR error rates answer different questions, and several category pages remain early or forthcoming. Procurement teams should therefore inspect each task's sample, aggregation method, evaluator population, and release status rather than compare numbers across unrelated categories. ### The ASR benchmark predates the August launch Voice of India's established speech-to-text benchmark covers **15 Indian languages, 306,230 utterances, 536 hours of speech, and 36,691 speakers** across 139 regional clusters, according to the project's research paper. Its audio comes from unscripted telephonic conversations rather than clean read speech. The benchmark uses multiple accepted lexical and spelling variants so that code-mixed or non-standardized Indian speech is not penalized solely for orthography. Its private held-out set is designed to limit test contamination, while a public sample supports debugging. The published analysis also reports performance by geography, audio quality, speaking rate, gender, device, and age instead of relying only on a national average. For teams building or buying speech and multimodal systems in India, the platform's value is its emphasis on subgroup and deployment-specific evidence. Its usefulness will depend on transparent task protocols, stable test-set governance, and clear separation between live benchmark results and placeholder or forthcoming categories. ## Key Points - 1The August 4 launch broadens Voice of India from its earlier ASR benchmark into a multi-category evaluation platform for Indian deployment conditions. - 2The mature speech benchmark covers 15 languages, 306,230 utterances, 536 hours, 36,691 speakers, and 139 regional clusters. - 3Teams should evaluate each category's protocol and release status separately because human-preference rankings, OI-WER scores, and forthcoming tasks are not interchangeable. ## Scoring Rationale The expanded platform addresses a practical evaluation gap for Indian languages and deployment contexts, anchored by a substantial ASR benchmark, while several broader categories remain early or forthcoming. ## Sources Primary source and supporting public references used for this report. ## View 3 more sources Practice interview problems based on real data 1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with. [Try 250 free problems](/problems)