What if I told you… the limits of your language are the limits of your thinking? Not metaphorically. Literally. Pehaps, humans are the only species on Earth that can turn invisible thoughts into such complex structured sound waves… and have another brain reconstruct them into meaning. Now here’s where it gets even more fascinating. The Pirahã tribe in the Amazon has no words for exact numbers—so they don’t think in precise quantities the way most of us do. In German, a word like Schadenfreude captures an emotion that takes a sentence in English. In Urdu, words like “Hijr” and “Visaal” hold the pain of separation and the joy of reunion—entire emotional journeys in a single word. In Telugu, words like “Mamakaram”, “Atmiyata”, and “Paruvu” capture belonging, soul-level connection, and social dignity—ideas that shape decisions and identities, yet don’t translate cleanly into English. In Japanese, what is not said often carries more meaning than what is. And in Sanskrit, a phrase like “Tat Tvam Asi” (You are That) compresses an entire philosophy of existence into three words. 👉 Different language. Different thinking. Different reality. So ask yourself: Are you thinking your thoughts… or your language’s thoughts? If you learned 5 new languages, would you become 5 different thinkers? What thoughts are you incapable of having today? Now comes the uncomfortable part. There are about 7,000 languages in the world today. And we are losing one every two weeks. By 2100, we may lose 50–90% of all languages. Let that sink in. We are not just losing words. We are erasing entire cognitive universes. Now enter AI and Large Language Models. Machines that don’t “live”… but are trained on our language are beginning to: write reason persuade even simulate empathy So the real question is no longer: 👉 Can AI communicate like humans? It is: 👉 If language shapes thinking… what happens when machines master language better than humans? But here’s the gap—today. AI doesn’t experience. It doesn’t intend. It doesn’t understand meaning the way we do. It predicts. Brilliantly. At scale. At speed. But the gap is narrowing. Fast. And here’s the part we’re not talking about enough: 👉 AI will only “think” in the languages and data we preserve. If thousands of languages disappear… their ways of thinking disappear from AI too. We risk building incredibly powerful systems… trained on a shrinking slice of human cognition. So this is not just a technology conversation. It’s a civilization design decision.

🚨 Most AI speech models ignore half the world’s voices. Guess what just changed? Google AI just dropped WAXAL – a massive multilingual African speech dataset for ASR and TTS. 60+ languages, 1,000+ hours of real‑world recordings, all open‑source. 👉 Struggle: Your model stumbles on Swahili or Yoruba because there’s no data. WAXAL fixes that. 👉 Step 1 – Grab the dataset from the Google repo (free, no gate‑keeping). 👉 Step 2 – Fine‑tune your ASR/TTS model on the target language slices. 👉 Step 3 – Deploy and watch accuracy jump from “meh” to “wow” in real time. 🎯 Analogy: It’s like giving your AI a passport to every African market. 💡 Save‑worthy rule: If you don’t train on local data, you’ll always be biased – period. The second‑order effect? More inclusive products, new markets, and a louder African tech voice. What African language are you most pumped to hear your AI speak? Let’s discuss in the comments. If you’ve already experimented, share your results – the community learns together. #AI #SpeechDataset #WAXAL #ASR #TTS

🚨 Josh Talks has developed what it describes as the world’s first full-duplex conversational AI model in Hindi, a system that can listen and speak at the same time, enabling more natural voice interactions compared with traditional voice assistants. “This research shows that when models are trained on large-scale natural dialogue, they can begin to learn the rhythm of how people actually speak to each other. By building this infrastructure using Hindi conversations, we have demonstrated that full-duplex conversational systems can be developed beyond English,” said Shobhit Banga, co-founder of Josh Talks. Josh Talks Read to know more 👇 https://lnkd.in/gHfj4Dqp

Over the weekend, I started experimenting with WAXAL, the newly released open-access African speech dataset by Google Research, and focused my first fine-tuning run on Shona ASR. What caught my attention about WAXAL is that it is not just another speech benchmark. It is a serious attempt to close the speech-data gap for underrepresented African languages, with publicly released data for both Automatic Speech Recognition (ASR) Text-to-Speech (TTS). For my first pass, I narrowed the scope to the Shona (sna_asr) subset and set up a Colab-based fine-tuning pipeline using Meta’s Massively Multilingual Speech model (MMS) (facebook/mms-1b-all) as the base checkpoint. A few early observations from the work so far: The labeled Shona subset is already substantial: about 14,109 train, 1,727 validation, and 1,749 test examples. The audio-text pairs are rich and descriptive, which is great for realism, but it also makes the ASR task harder because the utterances are often long, natural, and less constrained than scripted speech. Even in early outputs, the model is clearly capturing the semantic structure of many utterances quite well. The biggest errors I am seeing are not random failure, but more interesting issues like: spelling and orthographic variation, word boundary inconsistencies, near-miss lexical substitutions, and normalization problems between reference transcripts and predictions. On a small manual error-analysis sample, one thing stood out immediately: The model is often “almost right” in a linguistically meaningful way, but standard exact-match evaluation can make those predictions look worse than they actually are. That makes me think the next important step is not only more training, but also better text normalization and evaluation design for Shona ASR. This is one of the reasons I find African language AI so interesting: the challenge is not just model training, but building systems that respect the realities of language variation, transcription practice, and low-resource data conditions. Next steps on my side: continue fine-tuning and evaluate full-run performance, improve transcript normalization, inspect recurring error patterns more systematically, and test how much performance can improve with cleaner preprocessing. It is still early, but this already feels like a meaningful direction: using open multilingual speech resources to build more practical ASR pipelines for African languages. #AI #MachineLearning #ASR #SpeechRecognition #NLP #LowResourceLanguages #AfricanLanguages #Shona #WAXAL #MMS #DeepLearning #DataScience #OpenSourceAI Learn more and check out the WAXAL dataset here: goo.gle/4cy4dqC

I have been testing Qwen3-TTS-12Hz-1.7B-CustomVoice model for Generative AI voice generation from text. It looks impressive with various personas/speakers, voice description and their own native language, plus the multi-lingual support is very promising. I was testing it with this short train announcement in Japanese: 次は中目黒です。ドアは右側です。(you will hear the English version at the end of the video). I intentionally used the proper Kanji, not Hiragana, and swapped the original '出口 は' with 'ドアは'. The first announcer tripped several times when pronouncing 中, which should be pronounced as 'なか (naka)'. Instead, it came out as 'chu' or even 'tachi', giving me "Tsugi wa CHU-meguro desu." The speaker was not 'wrong', it was just being linguistically loyal to her training distribution. The second announcer performed much better, though the first character 次 came out strange a few times. At one point it sounded like 'tsuku'. My suspicion: a slight vowel/phoneme drift, or 次 segmented in a way the model doesn’t strongly recognize as Japanese, falling back to safe phonemes like “tsu” or “tsuku”. In the second English version by the same announcer, I find it interesting that Meguro sounded like Megyuro. Maybe the uncle was hungry and wanted to go to a 牛肉 restaurant. Imagine the train stops, the doors open, and suddenly .... "Ohh .. it smells good!" 😄 Lastly, I’m glad the model couldn’t replicate the iconic voice of Donna Burke on the Shinkansen. Some things are better left to humans and shouldn’t be replaced by AI. All in all, it was a very enlightening experiment in Value Discovery. It highlights the surprising ways that cross-lingual TTS models interact with speaker personas, phonetic biases, and stochastic decoding. It also underscores the importance of language alignment, persona tuning, and careful scenario testing, and reinforces how stress-testing AI systems at the edges can uncover hidden failure modes, edge-case behaviors, and opportunities for refinement before deployment. Without careful tuning, calibration, and testing, using such models in a public-facing application could easily lead to embarrassing or confusing experiences. The model: https://lnkd.in/gDvsBgJt #MultilingualTTS #TextToSpeech #GenerativeAI #VoiceAI #ValueDiscovery

Intron Voice AI, a Lagos-based AI startup has officially launched #Sahara-v2, its flagship, second-generation voice AI model, effectively raising the bar for global speech recognition. The model supports 24 new African languages and mastered over 500 distinct African English accents, bringing the company’s total language coverage to an unprecedented 57. https://lnkd.in/dH3WmGhJ

If your Voice AI agent sounds smart… how do you actually know it’s working? Not in demo calls. Not in internal testing. In real, messy production. If you’ve ever searched: How to evaluate Voice AI agents How to make sure my Voice AI agent is working properly What metrics should I track for Voice AI This is your checklist. Here’s what actually fails in production — and what you should be measuring. 1️⃣ End-to-End Latency This is the time from: 👉 When the user stops speaking 👉 To when the bot starts responding Users don’t experience ASR, LLM, and TTS separately. They experience silence. ~300ms → feels natural ~700ms → feels hesitant ~1s+ → feels broken Measure total latency, not just individual components. 2️⃣ Task / Goal Completion Rate Did the agent actually complete what the user came to do? - Appointment booked - Issue resolved - Refund processed - Form submitted Benchmarks: 80–85% → early stage 90%+ → production-grade A call that “sounds good” but doesn’t complete the task is still a failed call. 3️⃣ Speech Recognition Accuracy (WER) Before reasoning starts, the system must hear correctly. Word Error Rate (WER): (Substitutions + Deletions + Insertions) / Total Words × 100 Rule of thumb: <10% → production-ready Higher than that → downstream errors stack fast If the agent hears wrong, everything after that is damage control. 4️⃣ Conversational Flow & Turn-Taking Track: - Interruption rate - Barge-in latency (how fast the bot stops when user speaks) - Overlap duration We’ve seen systems where transcripts looked perfect. But once measured, the bot was constantly speaking over users. Users rarely complain. They just stop using it. 5️⃣ Language Alignment (Critical in India) Understanding a language is not the same as staying aligned with it. In India, we mix naturally: “Haan that works.” “Meeting kal shift kar do.” Measure: - Did the agent continue in the user’s language? - Did it switch unnecessarily? - Did it drift back to a default language? Multilingual markets blend. Your evaluation framework should too. 6️⃣ Hallucination & Compliance Track: - Hallucination rate - Policy violations - Tool correctness Voice makes incorrect answers more dangerous because they’re delivered confidently and instantly. A fast wrong answer is worse than a slightly slower correct one. If you’re evaluating Voice AI agents, your framework should include: - End-to-End Latency - Task Completion Rate - WER (Speech Recognition Accuracy) - Turn-taking Quality - Language Alignment - Hallucination & Compliance At Noveum AI, we make sure our customers don’t operate with even one such blind spot. 💯

The biggest land grab is happening in Voice AI Agents right now! Everyone is building voice ai but it's a huge physics and engineering challenge. How do you ensure your voice AI agent actually works? What all you need to measure, evaluate and fix? Read below to find out what you need to do to make sure you voice agent is working properly:

At Nexenda, building AI that is truly inclusive and representative requires not just engineering depth but research that challenges assumptions, expands linguistic coverage, and embeds community participation at its core. Championing this work is Benjamin Ogbonna, a Member of Technical Staff in Research and Engineering whose contributions pushes the frontier of language processing with safety and humanity at the center. Benjamin is the creator of Nigeria’s first Accented Text-to-Speech (TTS) model, a natural and expressive voice synthesis system tailored for Nigerian English, designed to power narration, voice cloning, conversational assistants, and voice-enabled applications that sound like home. Ben’s work extends to foundational speech resource curation for Nigerian Common Voice datasets covering Hausa, Igbo, and Yoruba which he open-sourced to further unlock resources for home-grown Automatic Speech Recognition, accent recognition, and inclusive language systems, pushing open science forward in broader ASR, TTS, and NLP work for global researchers and developers. He is also the developer of AI Chat Wrapped, an LLM-powered analytics system that transforms AI chat histories into personalized insights, topic trends, and behavioral personas, inspired by the idea of “Wrapped” experiences but designed for meaningful reflection on AI interactions. This project shows his ability to blend research, product thinking, and analytics engineering into tools that help users understand and grow with AI Ben’s academic research includes peer-reviewed work such as Accent-Aware Text-to-Speech for Nigerian English, presented at NeurIPS 2025, which demonstrates how community-curated data and advanced architectures can produce expressive, accent-faithful speech models that perform well on real evaluation metrics. This blend of publication-grade research with practical engineering is central to how he delivers AI systems that are structurally sound but also culturally and linguistically relevant. Benjamin plays a pivotal role at Nexenda, not just in MLOps, but also in shaping research directions that ensures that our voice and language technologies are built on scientific rigor, inclusive data, and collaborative research practices. He partners with other Members of Technical Staff to build and accelerate high-impact work in line with our mission and long-term vision. Outside of professional work, Benjamin is an Ambassador at DeepLearning.AI where he mentors early-career engineers and promotes responsible AI through community engagement and training sessions, impacting 4,000+ participants globally. #AISafety #FrontierAI #AfricaTech #SafeAGI

India and the Next Frontier of AI: Linguistic Inclusion!! Much of the global conversation around AI focuses on bigger models, more compute, and faster capabilities. But one of the most important questions for the future of AI is far simpler: Who does AI actually understand? While most AI systems today work best in a handful of dominant languages, the world speaks over 7,000 languages ,many of which remain invisible to technology. This is where #India may quietly shape the next frontier of AI. With its extraordinary linguistic diversity, India is not just a user of AI technologies ,it is becoming a living laboratory for multilingual and inclusive AI systems. Initiatives like #ProjectAstitva, emerging within the ecosystem of BHASHINI - (Digital India BHASHINI Division), demonstrate what this future could look like. By working with tribal communities and language experts, the initiative is helping AI systems understand languages such as Bhili, Pawari, Kokani, and Mavchi — languages that have historically been outside the digital ecosystem. This represents a fundamental shift in how AI is built: • From global models to locally grounded intelligence • From data extraction to community participation • From technology access to digital inclusion In many ways, India’s linguistic complexity may turn out to be one of its greatest advantages in shaping the future of AI. Because the next phase of artificial intelligence will not be defined only by how powerful models become, but by how many voices they are able to understand. And if AI is truly to serve humanity, it must learn to listen to every language, every community, and every voice. Amitabh Nag Subrata Chatterji IndiaAI #AI #ArtificialIntelligence #AIIndia #InclusiveAI #LanguageAI #DigitalIndia #CognitiveAI