The best speech recognition models, according to hackers

25 speech recognition models recommended in Hacker News comments over the past year, ranked by how many separate people recommend each one.

All speech recognition models

OpenAI WhisperA free, open speech-to-text model people run locally for accurate transcription and dictation across many languages, though it tends to invent text during silence and garble names.164recommend
NVIDIA ParakeetA small, fast local speech-to-text model that runs well even on CPUs and old laptops, though some find Whisper more accurate and note stutters, repeated words or dropped text.144recommend
Mistral VoxtralPeople recommend Voxtral for fast, accurate transcription of jargon, rap and mixed-language speech at a low price. Critics most often report wrong-language output, such as Ukrainian turned into Russian.69recommend
Moonshine open-weights STT modelsOpen-weights speech-to-text models praised for fast streaming latency and accuracy close to larger models, running on small hardware. Critics say Parakeet beats them on accuracy and speed.37recommend
whisper.cppFast, free, fully local speech-to-text that people run for dictation, subtitles and meetings. Critics say speed depends heavily on hardware, especially without a high-end GPU.33recommend
Apple SpeechAnalyzer APIRecommended as a fast, accurate on-device speech engine that ships free with macOS; critics note limited language support and missing language autodetection.25recommend
Transcribe.cppFree, fast, fully local transcription library with an embeddable C++ engine, language bindings and streaming support, praised for accuracy and easier integration than its predecessor.21recommend
Voxtral Mini 4B RealtimeOpen-weights speech-to-text model under Apache 2.0 that people run locally for real-time dictation. The most common caveat is its size: a 2.5GB download and memory footprint, and it is too slow on some machines.17recommend
Qwen3-ASR 1.7BPeople recommend this open speech-to-text model for accurate transcription that runs fast on local hardware, including laptops and Apple Silicon. Critics say it is only good for English.15recommend
Whisper Large v3 TurboRecommended for accurate, fast speech-to-text on accents, jargon and several languages, including real-time use on Apple silicon.12recommend
faster-whisperRecommended as a fast, free way to run Whisper speech-to-text locally, even on CPU. Main caveat: its output needs getting used to.11recommend
Cohere TranscribeOpen-source Apache 2.0 speech-to-text model that people find accurate, fast and competitive on benchmarks. Critics say its quality varies by task and it lacks custom vocabulary support.11recommend
WhisperXOpen-source CLI for fast transcription and speaker diarization that people run locally; setup is fiddly and speaker labels can be inconsistent.10recommend
Gemini-3.5-TranscribePeople recommend it for high transcription accuracy and good formatting, with latency that is still a bit slow. Critics note it can simplify precise wording and cut off the start of audio.10recommend
SonioxPeople recommend Soniox for accurate transcription that beats rivals like Deepgram, with quick setup and lower prices than many competitors. The recurring caveat is that it is a bit expensive.9recommend
Omnilingual ASR by MetaFree, openly downloadable speech recognition for many languages, praised for rare-language coverage. Critics doubt accuracy outside major languages and say the demo misheard many words.9recommend
DeepgramPeople recommend Deepgram for accurate speech-to-text, diarization and low latency, with a free tier for personal use. Its text-to-speech is seen as weaker than some rivals.8recommend
FluidAudioFree Swift library and CLI for running Parakeet speech recognition on Apple devices, with fast inference. Critics note it lacks streaming support.7recommend
Mistral Voxtral MiniPeople recommend it for accurate speech-to-text that is also cheap to run, either locally or through the API. One critic noted that the API price rose from 0.003 to 0.0035 dollars.7recommend
WhisperKitDevelopers recommend it as a free, fast, offline speech-to-text toolkit for Apple Silicon Macs, with support for speaker diarization and concurrent inference.6recommend
ElevenLabs Scribe v2Recommended for accuracy on non-English speech, including French dialogue and European languages. One mixed comment says its European-language support is only half decent.6recommend
Nvidia Parakeet Unified EN 0.6BPeople recommend it for English speech-to-text, praising its offline accuracy compared with earlier Parakeet versions and its streaming output. No recurring criticism appeared.6recommend
sherpa-onnxRecommended for offline speech recognition and text-to-speech on local hardware, with bindings for many languages and streaming support on Linux. Caveat: larger memory footprint for higher accuracy.5recommend
SpeechmaticsPeople recommend it for speech recognition accuracy, including multi-speaker audio, and its wide language range. The caveat is that it is a bit more expensive, at $0.004/min.5recommend
AssemblyAIPeople recommend it for accurate transcription, good cost-performance and clear data terms. One critic reports rival services beating it on streaming benchmarks.5recommend