Transcription models.

Deepgram Base model — high volume, cost-effective

Google's latest generative ASR foundation model — 85+ languages

Optimized for short queries and voice commands

General-purpose model

Deepgram Enhanced model — high accuracy for uncommon words

Highest accuracy model — 55+ languages, best-in-class

First conversational ASR model built for voice agents — model-integrated endpointing

Google Gemini batch transcription with word timestamps, optional speaker labels and custom vocabulary.

GPT-4o Mini optimized for fast transcription

GPT-4o optimized for transcription with improved WER

GPT-4o transcription with built-in speaker diarization

OpenAI GPT Transcribe — OpenAI's current file-transcription model and the successor to GPT-4o Transcribe and Whisper. Detects the spoken language automatically and accepts custom vocabulary. Returns the transcript text without word-level timestamps.

xAI Grok Voice Transcribe 2.0 batch speech-to-text with word timestamps, speaker diarization, and automatic language detection.

Conformer model for long-form audio (minutes to hours)

Conformer model for short utterances (< 60s)

Microsoft MAI-Transcribe-2 — Microsoft AI's in-house transcription model, in public preview on Azure Speech. 60 languages with automatic language detection and mid-sentence code-switching, speaker diarization, word-level timestamps and custom vocabulary. Accepts WAV, MP3 and FLAC.

Meta Muse Voice Transcribe for batch WAV audio with turn timestamps, speaker diarization and language detection.

Deepgram's Nova-2 speech recognition

Optimized for human-to-bot interactions (IVR, voice assistants)

Optimized for earnings calls with finance vocabulary

Optimized for conference room audio

Optimized for low-bandwidth phone call audio

Optimized for low-bandwidth single speaker voicemail

Deepgram's flagship model — 53% lower WER vs competitors, code-switching across 55 languages

Alibaba Qwen3-ASR Flash — multilingual batch transcription (25+ languages) with word-level timestamps and automatic language detection, served from Model Studio (Singapore). Handles files up to 12 hours.

Rev AI Reverb — async speech-to-text built on Rev's Reverb ASR model (trained on 3M+ hours of human-transcribed audio), with speaker diarization for up to 8 speakers and word-level timestamps. English-core with broad language support via the async API.

State-of-the-art batch STT — 90+ languages, speaker diarization, audio tagging

Most accurate low-latency STT — <150ms, 90+ languages

Gladia Solaria-1 — EU-hosted real-time streaming speech-to-text across 100+ languages with native code-switching and ~103ms partial latency. Word-level timestamps and automatic language identification; speaker diarization is available for batch (Solaria-3), not this streaming model.

Gladia Solaria-3 — EU-hosted batch speech-to-text tuned for the most accurate business audio across five core European languages (English, French, German, Spanish, Italian), with speaker diarization, custom vocabulary, and word-level timestamps.

Azure default STT model — 140+ languages, diarization, word timestamps

Cost-effective model — fast turnaround, 55+ languages

Soniox async batch transcription — one multilingual model across 60+ languages with speaker diarization, per-word timestamps, and automatic language identification, at a low flat token-based rate.

Soniox realtime streaming transcription — low-latency multilingual STT across 60+ languages with speaker diarization, per-word timestamps, and automatic language identification over a single WebSocket.

Optimized for telephony audio (8kHz)

AWS foundation model-powered ASR — 100+ languages

Medical transcription with HIPAA eligibility

AssemblyAI Universal-3.5 Pro Realtime — the flagship streaming model, with native mid-sentence code-switching across 18 languages and conversation context carried across turns. Word-level timestamps and stable partials that are not rewritten as they arrive; speaker diarization is available on the batch model, not this streaming one.

AssemblyAI's most powerful speech language model — up to 1000 keyterm phrases

Purpose-built for real-time voice agents — ~300ms immutable transcripts

Multilingual streaming STT — English, Spanish, French, German, Italian, Portuguese

Mistral Voxtral Mini Transcribe — EU-hosted batch transcription with speaker diarization, segment timestamps, and custom vocabulary (context bias). Strong accuracy at very low cost; handles recordings up to 3 hours.

OpenAI's Whisper API model

OpenAI Whisper large-v3 served on Groq's LPU hardware — top Whisper accuracy at Groq speed and cost, with word-level timestamps and automatic language detection.

OpenAI's Whisper large-v3 model

OpenAI Whisper large-v3-turbo hosted by Cloudflare Workers AI, with word-level timestamps and automatic language detection.

OpenAI Whisper large-v3-turbo served on Groq's LPU hardware — very fast, very low cost batch transcription with word-level timestamps and automatic language detection.