In the early 2010s, deep nets replaced key parts of HMM-GMM systems and sharply cut speech error rates.
For decades, speech recognition relied on HMMs with Gaussian mixtures. Around 2010–2012, teams at Toronto, Microsoft, Google, and IBM showed DNNs substantially improved acoustics.
The shift was among the first mass industrial deployments of deep learning, even before AlexNet’s fame among the broader technical public.
Today speech is dominated by end-to-end and multimodal models, but the HMM→DNN bridge was a critical rung of enterprise trust in deep nets.
Around September 11, 2026, Microsoft launched MAI-Transcribe-2 with claims of leadership across about 60 languages and indicative pricing near $0.10 per audio hour. It is positioned against prior MAI generations and OpenAI transcription APIs. Source: Trew Knowledge AI This Week for September 11.
xAI introduced Grok Voice Transcribe 2.0 as its most accurate speech-to-text model, claiming twice the accuracy of version 1.0 at the same $0.10/hour batch and $0.20/hour streaming price, per the lab’s announcement.
On 3 September 2026, Microsoft AI introduced MAI-Transcribe-2 as the fastest, most accurate and cheapest speech recognition versus rivals such as GPT-Transcribe, Gemini 3.5 Transcribe and ElevenLabs Scribe v2, at $0.10 per hour as a launch offer.