STT / Speech-to-Text

Comparison of speech recognition solutions for conversational pipelines (2025–2026). Benchmarks, strategic stakes, and decision questions.

🎯

Strategic Framing — STT: Much More Than a WER

The Real Question

WER doesn't tell the whole story. The real question is: where is your voice data processed? Who has access to it? What is your strategy if the provider raises prices or gets acquired?

Infrastructure Spectrum

Cloud-only (Google, AssemblyAI) → Cloud+VPC (Deepgram, Azure) → On-premise (Azure containers, Inworld) → Self-hosted open-source (Whisper, Voxtral). Each step increases sovereignty and reduces lock-in.

2026 Market Signal

Deepgram ($1.3B) and AssemblyAI ($158M) are acquisition targets. Voxtral and Whisper are closing the quality gap. Inworld STT raised prices 400%+. The market is consolidating fast.

Questions to ask before choosing: Is your users' voice data subject to GDPR or Swiss nLPD? Do you need diarization or PII redaction? Does your architecture allow switching to a self-hosted model without major rework? What is your acceptable WER threshold for your specific domain?

Real-time cloud STT APIs — Deepgram Nova-3 is the latency reference (75ms TTFA). Whisper large-v3 is the open-source quality standard. AssemblyAI Universal-3 Pro leads the multilingual WER benchmark — Voice Agent API $4.50/hr (launched Apr 29, 2026).

Click a header to sort

SolutionTTFA?WER?Streaming?MultilingualDiarization?Price/hrSovereign?
Deepgram Nova-3
Phase 1 MVP — ASR streaming
75ms7.2%36 langs$0.216/hr
Inworld STT
Phase 1 MVP — STT émotionnel + Axe 2 Avatar Behavior
92ms5%100 langs$0.36/hr
AssemblyAI Universal-3 Pro
Voice Agent Pipeline — Référence précision
150ms4.9%99 langs$0.372/hr
Azure Speech (Microsoft)
Souveraineté suisse — institutionnel
180ms5.9%100 langs$1/hr
Google Speech-to-Text v2
Multilingue — option secondaire
200ms2.7%125 langs$0.36/hr
Deepgram Nova-3
Phase 1 MVP — ASR streaming
Latency75ms
10
Accuracy7.2% WER
8
Price access7/10
7

Deepgram Nova-3 is the industry reference for real-time voice agent ASR. 75ms P90 streaming latency, 36 languages, built-in VAD and endpointing. Voice Agent API at $4.50/hr provides a complete STT+LLM+TTS pipeline in a single WebSocket, with BYO LLM (GPT-4, Claude, Gemini) and function calling for RAG integration. No native voice cloning — Aura-2 TTS offers 36 preset voices; external TTS (ElevenLabs, Cartesia) can be integrated. On-premise deployment available for Enterprise (partial sovereignty). EU endpoint available. Used by Tavus, Simli, and most commercial avatar platforms.

Full details →
Inworld STT
Phase 1 MVP — STT émotionnel + Axe 2 Avatar Behavior
Latency92ms
10
Accuracy5% WER
9
Price access7/10
7

Inworld STT (2025–2026) is the most feature-rich cloud STT API for interactive voice agents. Sub-100ms documented latency, 100+ languages via multi-provider routing (Whisper large-v3 + AssemblyAI). Unique real-time voice profiling extracts emotion (happy/calm/angry/frustrated), accent, age, pitch, and vocal style on every streaming chunk. Realtime API (full pipeline STT+LLM+TTS) from $0.015/min — 4x cheaper than OpenAI Realtime ($0.06/min). Native voice cloning: built-in + cloned + custom voices (up to 3,000 custom voices on Growth plan). RAG via function calling (tool calling mid-conversation). ZDR support. On-premise available on Enterprise. Drop-in compatible with OpenAI Realtime API.

Full details →
AssemblyAI Universal-3 Pro
Voice Agent Pipeline — Référence précision
Latency150ms
7
Accuracy4.9% WER
10
Price access5/10
5

AssemblyAI a lancé sa Voice Agent API le 29 avril 2026 : pipeline complet STT+LLM+TTS en une seule connexion WebSocket à $4.50/hr flat. Universal-3 Pro Streaming (u3-rt-pro) est le modèle STT temps réel, avec turn detection sémantique+acoustique, barge-in natif et session resumption. Tool calling (JSON Schema) permet d'intégrer un RAG custom (Pinecone, LlamaIndex, etc.) via function calling — pas de RAG natif mais intégration externe complète. Pas de clonage vocal natif : les voix disponibles sont prédéfinies (18+ voix EN/multilingual), mais il est possible d'intégrer un TTS externe (ElevenLabs, Cartesia) pour un clone vocal custom. LeMUR AI features (summarization, Q&A, sentiment) disponibles sur transcriptions. 99 langues, diarisation, timestamps mot-à-mot.

Full details →
Azure Speech (Microsoft)
Souveraineté suisse — institutionnel
Latency180ms
7
Accuracy5.9% WER
9
Price access3/10
3

Azure Speech offers enterprise-grade STT with 100+ languages, custom model training, and Swiss data center (Zurich). 5.9% WER on English. Disconnected container deployment for partial sovereignty. Most expensive option but strongest enterprise compliance. Swiss German custom model available.

Full details →
Google Speech-to-Text v2
Multilingue — option secondaire
Latency200ms
6
Accuracy2.7% WER
8
Price access6/10
6

Google Speech-to-Text v2 avec Chirp 3 (GA mai 2026) : 2.7% WER EN (Artificial Analysis), 125+ langues, streaming gRPC bidirectionnel. MedASR open-source lancé fin 2025 (domaine médical). Expansion offline majeure (avril 2026). Pas de clonage vocal natif (via Google TTS séparé). RAG via tool calling + Vertex AI Search. Cloud uniquement, pas d'on-premise.

Full details →