August 2024 Summaries

In the post-pandemic economy, voice AI is presented as a tool helping non-technology companies adapt to changing customer expectations, labor shortages, and operational pressures. Macy’s illustrates its use in retail, where voice-enabled personalization, recommendations, advertising, ordering, and customer analytics can connect online and in-store experiences as retailers respond to shifts in e-commerce demand. The technology, which has evolved from early speech synthesis to advanced generative systems and voice assistants, is also being adopted in healthcare to automate administrative work and support diagnostics, in education to reduce teachers’ paperwork and provide customized learning assistance, and in financial services to improve customer support, transactions, and fraud prevention. Hospitality businesses, including hotels, restaurants, and bars, use voice AI for personalized guest services, reservations, orders, housekeeping, and routine staff tasks. Examples such as Intermountain Health, ISTE’s Stretch AI, Deutsche Bank, Macy’s, and Marriott suggest that organizations view voice AI as a way to improve accessibility, personalize services, streamline workflows, and allow employees to focus on more complex or customer-facing work.

Voice AI has progressed from early demonstrations such as IBM’s 1961 performance of “Daisy Bell” to multimodal systems like OpenAI’s GPT-4o, which can process text, audio, images, and video while responding in near-real time. The examples described range from practical applications, including virtual assistants, financial fraud detection, healthcare scheduling, restaurant ordering systems, and educational tutoring, to creative and experimental uses such as cloned-voice podcasts, AI-assisted music production, and the restoration of John Lennon’s vocals for the Beatles song “Now and Then.” Demonstrations from ChatGPT, Kyutai’s Moshi chatbot, and conversational AI systems highlight increasingly natural speech, fast response times, and humanlike interaction patterns, while also raising questions about model behavior, authenticity, artistic ethics, copyright, and the boundaries between simulated and human communication.

Meta’s Llama 3.1 and Mistral’s NVIDIA-developed Mistral NeMo are new openly available language-model families that reflect growing competition in capable, locally deployable AI systems. Released in July, Llama 3.1 ranges from 8B to 405B parameters, offers instruction-tuned versions, a 128,000-token context window, multilingual capabilities, and broad Meta product integration, while Mistral NeMo is a 12B-parameter model positioned between smaller 8B and much larger 70B-class models, with the same context size and a Tekken tokenizer intended to improve compression for languages such as Korean and Arabic and for code. Official benchmark comparisons suggest Llama 3.1 8B outperformed NeMo on the shared MMLU metric, while Meta’s human evaluations positioned the 405B model competitively against leading proprietary systems, though its full-precision hardware requirements remain substantial. Informal local tests found that Llama 3.1 8B more reliably rejected an invented factual premise and correctly handled a spelling question, whereas Mistral NeMo produced a functional two-player JavaScript Pong game more consistently despite sometimes hallucinating details around the false scenario. The discussion also notes that benchmarks and isolated tests cannot fully represent real-world performance, and contrasts open-model advances with proprietary developments such as GPT-4o’s emerging conversational voice mode.

Deepgram describes its Enterprise Voice AI Accelerator Program as an initiative to help companies move voice AI applications from prototypes to production, drawing on its experience operating voice models at scale. It identifies common adoption barriers, including uncertain return on investment, performance and latency limitations, talent shortages, data privacy requirements, and concerns about rapidly changing technology. The program offers technical support for architecture, model selection, customization, scaling, security, and deployment, along with early product access, developer credits, testing environments, and workshop opportunities. Deepgram also announced a hands-on Voice AI Agent Workshop with Groq, covering speech-to-text, text-to-speech, agent design, real-world use cases, open-source code, credits, expert Q&A, and networking.