The best speech-to-text models are not interchangeable. A model that produces an excellent offline transcript can still break a voice agent through slow partials, unstable revisions, or delayed endpointing. A fast streaming model, conversely, can be the right operational call even when a larger batch model wins on clean-audio accuracy.

This production-focused comparison covers six prominent speech-to-text models in 2026. They are ordered for evaluation, not ranked as universal winners: Smallest AI Pulse, Deepgram Nova-3, AssemblyAI Universal-3.5 Pro, OpenAI transcription models, ElevenLabs Scribe v2, and Cartesia Ink-2 The best STT models for any given team depend on audio conditions, interaction style, languages, infrastructure, concurrency, and acceptable cost.

Production comparison matrix

Benchmark discipline: Published WER figures are not directly comparable unless the dataset, normalization, audio preprocessing, model version, prompting, and scoring code are identical. Use published benchmarks as directional evidence, then validate every model on the same representative production audio before making a decision.

1. Smallest AI Pulse for latency-sensitive voice agents

Pulse belongs on the shortlist when conversational turn-taking matters more than producing a polished document minutes later. The production question is not simply how fast inference runs. Developers should record time to first useful partial, partial revision frequency, endpoint detection delay, and time to a committed transcript. That complete sequence is what determines whether a real-time agent actually feels responsive.

Test Pulse with packet jitter, codec transcoding, interruptions, and 8 kHz audio before committing it to phone automation. The streaming speech-to-text in production checklist covers reconnects, duplicate segments, and dropped audio.

Pulse is also worth evaluating when STT sits beside TTS and agent orchestration in a single voice stack. That can simplify support boundaries, but integration alone does not prove accuracy. Compare Pulse and Pulse Pro using current model cards, then run separate test suites for clean microphones, accented speech, background noise, and telephony. Review Smallest.ai's model pricing against expected audio minutes, silence handling, and peak concurrency rather than a single demo-call estimate.

2. Deepgram Nova-3 for telephony and configurable streaming

Deepgram has a mature streaming interface and a strong fit for contact-center engineering. The practical advantage is control: model selection, endpointing parameters, interim results, multichannel handling, diarization options, and vocabulary features can be composed around a call workflow. Every additional switch also expands the test matrix.

Deepgram is especially practical when:

• Audio arrives as 8 kHz mu-law or other phone codecs and must remain streaming.

• The application needs explicit control over end-of-turn behavior.

• Contact-center vocabulary requires key terms, formatting, or channel separation.

• Teams need established API operations at substantial concurrent-session volume.

Do not carry microphone WER numbers over to telephone call evaluation. Narrowband audio strips high-frequency cues, while hold music, speakerphone echo, and carrier transcoding introduce a different error profile entirely. Compare vendors on the same call set and inspect deletions of short acknowledgements, names, numbers, and negations. Published real-world STT latency benchmarks are useful methodology references, but reproduce them from your own region before drawing conclusions.

3. AssemblyAI Universal-3.5 Pro for meetings and transcript workflows

Meeting and media products often need more than word recognition. Speaker labels, timestamps, formatting, asynchronous processing, and downstream transcript intelligence can deliver more product value than shaving milliseconds from first-partial latency. AssemblyAI is strongest in that broader transcript workflow.

Treat its streaming and asynchronous models as distinct systems during evaluation. Batch transcription can exploit future context; streaming inference must emit hypotheses without hearing the rest of the sentence. Measure diarization error separately from WER, because a transcript can contain the right words under the wrong speaker. For meeting audio, test overlapping speech, distant microphones, room reverberation, and participant names added through supported vocabulary controls.

AssemblyAI is a sensible candidate for live captions and meeting platforms, particularly where enrichment reduces backend work. For a voice agent, scrutinize finalization delay and endpoint controls before treating its general streaming performance as equivalent to turn-taking performance.

4. OpenAI models for batch and realtime transcription

OpenAI spans two distinct deployment categories: hosted transcription models and open-source Whisper checkpoints that teams operate themselves. Collapsing both into a single label obscures major differences in inference latency, hardware requirements, batching behavior, streaming support, privacy posture, and cost.

Hosted OpenAI transcription is compelling for multilingual batch audio and teams already integrated into its API ecosystem. Whisper remains a practical self-hosted baseline where data control outweighs operational simplicity. It is not inherently a streaming model. Chunking it behind a WebSocket produces pseudo-streaming, with accuracy and boundary artifacts determined by window size and overlap.

Northflank's 2026 review reported Canary Qwen 2.5B at 5.63% average WER on the Open ASR Leaderboard. That independently reported result makes Canary relevant to self-hosted English evaluation, but it cannot be compared directly with vendor WER from different corpora.

For multilingual applications, build per-language and mixed-language test sets. Average multilingual scores mask weak locales. Code-switching tests should include mid-sentence language shifts, proper nouns from a second language, and scripts that require different normalization rules.

5. ElevenLabs Scribe v2 Realtime for multilingual realtime experiences

Scribe is relevant to developers pairing recognition with generated speech, captions, or multilingual media. ElevenLabs has promoted sub-150 ms latency for Scribe v2 Realtime. Treat that as a vendor-reported metric until the measurement boundary is reproduced on target audio. It may represent first token arrival rather than a stable partial or finalized utterance.

Production testing should score accent groups independently and include code-switching, crosstalk, music, and names. Inspect how aggressively partial words change, a caption system tolerates revisions differently from an agent that feeds partial text to an LLM. Confirm diarization, vocabulary controls, telephony support, retention policies, regions, and concurrency limits in the currently purchased plan rather than assuming the broader ElevenLabs platform applies to Scribe.

6. Cartesia Ink-2 for conversational streaming

Cartesia Ink is the specialist candidate in this group. Aimed squarely at realtime interaction, it deserves testing in environments where every stage of the voice loop carries a strict latency budget. That means measuring from audio capture on the client, not from the moment a server request reaches the model.

Ink's production fit depends on more than a fast median. Capture p50, p95, and p99 partial latency under peak concurrency, alongside disconnect frequency, recovery semantics, transcript duplication, and finalization delay. A younger service can perform well in benchmarks while still requiring more buyer-side validation around quotas, regional capacity, custom vocabulary, diarization, and language coverage.

For conversational systems, compare Ink with Pulse, Deepgram, and Scribe using the same endpointing policy. Without that constraint, a test measures different silence thresholds rather than model speed. A practical voice latency budget is outlined in designing voice assistants with STT.

Production verdict: choose by workload

Select an STT model from workload constraints, then validate it with representative audio.

Actionable shortlist:

• Real-time voice agents: Start with Smallest AI Pulse, Deepgram Flux, Cartesia Ink-2, and ElevenLabs Scribe v2 Realtime. Prioritize endpoint stability and final transcript latency.

• Phone calls and contact centers: Start with Deepgram and Pulse, then test 8 kHz codecs, crosstalk, names, numbers, and negation-heavy utterances.

• Live captions: Compare AssemblyAI, ElevenLabs, and Deepgram for partial stability, punctuation, delay, and accessibility requirements.

• Meetings: Favor AssemblyAI or a capable batch pipeline when diarization, timestamps, and overlapping speech matter more than turn latency.

• Batch transcription: Compare OpenAI's hosted models, self-hosted Whisper or Canary, and asynchronous commercial APIs on complete-file WER and throughput.

• Multilingual products: Shortlist OpenAI and ElevenLabs, then add regional specialists if per-language tests expose gaps.

• Self-hosted environments: Benchmark Whisper and Canary variants on owned hardware. Include GPU utilization, queue delay, autoscaling, observability, and engineering labor in cost.

A rigorous production evaluation requires a frozen audio set, one normalization script, identical scoring, and separate slices for language, accent, device, noise, and channel. Measure partial transcript latency, final transcript latency, endpoint false cuts, real-time factor, error rates, uptime, and cost per successful workload. Load tests must hold stream duration and concurrency constant across every candidate.

The best speech-to-text models are those that meet a specific application's quality and latency budget while remaining operable at peak load. Run a shadow deployment before migration, preserve raw audio where policy permits, and review errors by business impact rather than WER alone.