Which speech to text should you pair with an LLM?

There is no single best speech to text service to pair with a large language model (LLM). The right one depends on what the speech is for. Live speech that an LLM must answer or edit needs a streaming recognizer with a fast final transcript. Recordings need a batch model, and audio that must stay on your device needs an on-device recognizer. Benchmarks rank services on accuracy and speed, but the fastest and the most accurate service can be different ones, so the choice comes down to your job and a test on your own speech.

Choose by job first

Live speech: the fastest and the most accurate often differ

For a voice agent or live dictation, the delay of the final transcript matters as much as its accuracy. In Artificial Analysis's June 2026 streaming benchmark, Deepgram Flux was the fastest service on both the first partial and the first final transcript, at 0.020 seconds after end of speech for the final transcript, with a 7.36 percent word error rate ³. ElevenLabs Scribe v2 Realtime reached a 3.65 percent word error rate on its first partial transcript, with a latency of 0.13 seconds ⁴. A service that wins on speed does not always win on accuracy, so test both.

Partial text also changes. In the same benchmark, OpenAI's hosted GPT Realtime Whisper went from a 7.5 percent word error rate on its first partial transcript to 5.1 percent on the final transcript ⁵. Partial text is a draft that the recognizer revises. An LLM step should therefore run on the final transcript, which means waiting for it, and that wait is part of the cost.

Delay also grows when an LLM is a separate stage. The VoiceAgentBench study measured a pipeline of Whisper large v3 followed by the Qwen3 8B language model at about 800 milliseconds for the 90th percentile time to first token ⁶. An end-to-end speech model, Qwen2.5-Omni 7B, had about 40 milliseconds on the same hardware ⁷. An LLM placed after a recognizer adds its own wait on top of the recognizer's.

Recorded files: use a batch model and check the price

For a recording, you do not need a streaming service. A batch model can process the whole file. OpenAI's list price for gpt-transcribe is $0.0045 per minute, and gpt-4o-transcribe is $0.006 per minute ⁸. Prices change, so check the list on the day you decide. The guide to AI dictation cost explains how per-minute prices add up over a month.

A batch transcript is also easy to keep as the record. You can read the whole text against the recording before any LLM edits it.

Private audio: keep the recognizer on your device

If the audio is sensitive, a cloud service means your speech leaves your computer. An on-device recognizer keeps it local. Apple says transcription with SpeechAnalyzer is entirely on the device, although the language models must first be fetched ². Yap, which we make, is an on-device dictation app for Mac and Windows, so the per-minute cloud prices above do not apply to it. The cloud benchmarks in this article do not measure on-device engines, so test those separately.

It depends on five factors

• Latency: how long until the final text appears. Measure it in your app.
• Accuracy on your voice: benchmark audio is not your room, your microphone or your vocabulary.
• Language: a benchmark in one language may not carry over to yours, so test in the language you speak.
• Price: streaming and batch prices differ, and a cleanup pass by an LLM adds its own token cost.
• Privacy: cloud processing sends audio out of your device, and on-device engines do not.

What to do

1. Name the job. Decide whether the text is live, recorded or private. That choice removes most options.
2. Shortlist two services. For live use, pick two streaming services. For recordings, pick one batch model.
3. Test on your own speech. Record 20 to 30 sentences in your room, with your microphone, and score each service against what you said.
4. Measure the delay in your setup. Time the gap from end of speech to final text, including the LLM step if there is one.
5. Send the final transcript to the LLM. Keep the raw transcript, and use the LLM output as an edit you can reject.

For a closer look at how a model that hears audio compares with a recognizer followed by an LLM, read do LLMs make speech recognition more accurate. For the case of an LLM that listens directly, see is Whisper an LLM. For agents that answer rather than type, the delay is part of the product, and the guide to AI voice agents covers it. Deciding when a speaker has finished is covered by turn detection, and engines that run locally are described in on-device speech recognition. Our overview of the subject is speech to text and LLMs, within Voice and AI.