Automatic Speech Recognition (ASR): how speech-to-text models work and which one to use
Blog post from Gladia
Automatic speech recognition, or speech-to-text, converts spoken audio into written language by combining an audio encoder that interprets sound with a language model that generates coherent text. Modern systems are generally grouped into encoder-decoder, CTC, encoder-transducer, continuous-input speech LLM, and discrete-input speech LLM architectures, each differing in how audio and language components interact and in their trade-offs between accuracy, latency, streaming capability, scalability, and flexibility. Models such as Wav2Vec2, Whisper, Kyutai-STT, and NVIDIA’s Nemotron-Speech-Streaming illustrate these approaches, from self-supervised CTC systems and robust large-scale encoder-decoders to low-latency streaming designs. The discussion emphasizes that no model is universally best: organizations should evaluate candidates using their own audio, languages, accents, noise conditions, real-time requirements, deployment preferences, and total operating costs rather than relying solely on public benchmark scores such as word error rate.
| Trend | Post Mentions | Total Month Mentions | Posts | Companies | MoM |
|---|---|---|---|---|---|
| LLM | 37 | 6,051 | 1,178 | 246 | +14% |
| Vector Search | 19 | 1,674 | 383 | 126 | -31% |
| Real-time | 13 | 4,863 | 1,093 | 243 | +4% |
| Voice AI | 2 | 3,420 | 309 | 54 | +20% |
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.