Choosing an open speech‑recognition model today is less about the single WER number on the public leaderboard and more about fitting the model to your real‑world constraints. Start by checking the license: if you cannot afford attribution, drop any CC‑BY‑4.0 models (Parakeet, Canary‑Qwen‑2.5B, Kyutai STT) and stay with Apache 2.0 or MIT options. Next, verify language coverage – Cohere Transcribe covers 14 languages but has no auto‑detection, Granite Speech 4.1 handles six languages plus translation, while Qwen3‑ASR covers 52 languages and dialects. If you need low‑latency streaming, look at Voxtral Mini Realtime or Kyutai STT; batch‑only models like Parakeet TDT or Granite‑NAR will not meet real‑time turn‑taking needs without extra engineering.
Once you have a shortlist that satisfies license, language, and streaming requirements, run your own evaluation on a sample of your actual audio – accents, background noise, domain‑specific jargon. The spread between top models on the public benchmark is under one WER point, but on your data the differences will be larger and may re‑rank the models. Measure word error rate on your set, not the leaderboard number.
Finally, compute cost per audio‑hour on your own hardware. RTFx values quoted in model cards are obtained at large batch sizes on datacenter GPUs and do not translate directly to your deployment scenario. Profile inference latency and throughput with the batch size and concurrency you plan to use, then derive the effective cost.
By following this order – license → language → streaming → personal WER → cost – you avoid being misled by a single leaderboard score and pick a model that truly works for your product.
#AI #Product #SpeechRecognition #OpenSource #MLOps #VoiceTech