Microsoft AI's newest batch speech-to-text model (September 2026, in preview on Azure), covering 60 languages with automatic language detection, code-switching, speaker diarization and word timestamps. A real-time sibling, MAI-Transcribe-2-Streaming, followed on 2026-10-01.
Google's dedicated speech-to-text model, launched August 2026 and improving on Chirp 3, with auto-detection of 85+ languages, word-level timestamps, speaker diarization and custom vocabulary. Takes up to an hour of audio per request; a live variant streams.
Sarvam's newest speech recognition model, introduced August 18, 2026 and now its default recommendation. It keeps Saaras v3's five output modes and 22 Indian languages, adds Global English accents, and supports keyterm prompting.
Saaras v4 on Sarvam's asynchronous Batch API, for recordings up to two hours with speaker diarization and up to 50 prompted keyterms. It adds Global English recognition over the Saaras v3 batch model.
Qwen-Audio 3.0 short-audio recognizer (July 2026) for clips up to 5 minutes over HTTP, with hotwords, prompt context, 30 languages and seven Chinese dialect groups. Superseded by qwen-audio-3.1-asr-flash, which adds speaker diarization.
Qwen-Audio 3.0 long-file transcription (July 2026): asynchronous jobs on files up to 12 hours or 2 GB with speaker diarization, hotwords and context. qwen-audio-3.1-asr-flash-filetrans is now Alibaba's recommended offline model.
Real-time WebSocket recognizer from Qwen-Audio 3.0 for continuous transcription such as meetings, with unlimited stream length, hotwords and prompt context across 30 languages. Replaced as the recommendation by qwen-audio-3.1-asr-flash-streaming.
OpenAI's current high-accuracy speech-to-text model (July 2026) for files and committed Realtime turns, accepting context prompts, keyword hints and multiple language hints, at $0.0045 per minute. Named successor to whisper-1 and the GPT-4o transcribe models.
AssemblyAI's Sync API runs Universal-3.5 Pro on clips up to two minutes and returns the finished transcript in the same HTTP response, about 134 ms at p50 by AssemblyAI's figures. It launched in July 2026 for dictation and push-to-talk.
Universal-3.5 Pro is AssemblyAI's current flagship pre-recorded model, released July 2026, transcribing 18 languages with mid-sentence code switching. It predicts speaker labels jointly with the words and accepts a prompt describing the audio's domain.
Streaming version of Universal-3.5 Pro, released June 2026, that uses the voice agent's last question and earlier turns as context and offers min_latency, balanced and max_accuracy modes. Universal-3.6 Pro Realtime superseded it in September 2026.
Fun-ASR Flash snapshot from June 2026 for short recordings: HTTP transcription of clips up to 5 minutes with prompt context, covering 30 languages and the seven Chinese dialect groups. The newest Fun-ASR file model.
Melia 1 is Speechmatics' multilingual model, launched June 2026, transcribing speech that switches among 55+ languages in one pass and labelling each word's language. It requires language set to multi and lacks custom dictionary, confidence scores and translation.
Live-audio flagship from Soniox, released June 16, 2026, streaming transcription in 60+ languages with rebuilt speaker separation, faster semantic endpointing for voice agents and live translation across 3,600+ language pairs.
Soniox's current async model for recorded audio, released June 11, 2026, covering 60+ languages with reengineered speaker separation, better language identification and stronger context biasing for names, product terms and custom vocabulary.
Released June 2026, Solaria-3 is Gladia's newest model, limited to pre-recorded English, French, German, Spanish and Italian audio, one language per file. Gladia reports 26% lower English WER than Solaria-1 on real customer calls.
NVIDIA Nemotron 3.5 ASR (June 2026): a 600M-parameter cache-aware FastConformer-RNNT model that transcribes 40 language-locales from one checkpoint, built for low-latency streaming captions and voice agents as well as batch transcription.
NVIDIA's multilingual Nemotron ASR Streaming model, matching the 600M-parameter Nemotron 3.5 ASR checkpoint from June 2026: cache-aware streaming transcription across 40 language-locales (32 usable out of the box) with punctuation and automatic language detection.
NVIDIA Nemotron 3.5 ASR (June 2026), a 600M streaming speech recognizer covering 40 language-locales from one checkpoint via language-ID prompting, with automatic language detection, punctuation and selectable latency chunks; weights use OpenMDW-1.1.
June 2026 update to Microsoft AI's transcription model that widened coverage to 43 languages and added phrase-list keyword biasing and clean or verbatim output styles; file input, no diarization. Superseded by MAI-Transcribe-2 in September 2026.
