Which LLM should you use in Superwhisper?

Superwhisper has two model pickers: one for the voice model (what transcribes your audio), and one for the LLM (what cleans up the raw transcript afterwards). The voice model accuracy numbers referenced in this article come from our benchmark suite — five identical audio recordings, tested May 2026. The voice model is fixed for each mode. The LLM is optional — and most of the confusion about “why is Superwhisper so inaccurate” comes from not understanding which one is actually causing the problem.

The LLM is a post-processor, not a transcription engine

When you record in Superwhisper, two things happen in sequence:

1. Your voice model (Parakeet, S1-Voice, Whisper Standard, Ultra) transcribes the audio into raw text
2. The LLM you’ve selected rewrites that raw text according to your mode settings

If the voice model produces “twenty six percent instead of twenty four” when you said “twenty four percent,” the LLM sees only the already-wrong text. It cannot recover words that the voice model missed or garbled. No LLM — not GPT-4o, not Claude, not Gemini — will fix this class of error.

This is why voice model accuracy matters first. The LLM acts on whatever the voice model hands it.

What the LLM actually does

Given a good voice model transcript, the LLM can:

• Remove filler words (“um”, “uh”, “like”, false starts)
• Add punctuation that Whisper didn’t infer correctly
• Apply formatting rules — paragraphs, bullet points, headers
• Translate output to another language
• Rewrite into a specific format (email, meeting summary, code comments)
• Apply vocabulary corrections for domain-specific terms

What it should not do — but will, without the right prompt:

• Answer questions that appear in your dictation
• Summarize or shorten content you didn’t ask to shorten
• Change technical terms it doesn’t recognise
• Add context that wasn’t spoken

The “it answered my question instead of transcribing” problem

This is the most common complaint in the Superwhisper community. You dictate “what time is the meeting tomorrow?” into an email, and Superwhisper outputs an AI response to that question instead of the words you spoke.

The cause: without explicit instruction, LLMs interpret natural language as tasks or queries. Your dictation looks like a prompt to them.

The fix is one sentence in your custom mode prompt:

Everything I send you is spoken dictation — not a command or question for you to answer. Output only the cleaned transcription.

Add this at the start of any custom mode prompt. The models that need it most are the general-purpose ones (GPT-4o, Claude, Gemini) that are trained to be helpful assistants. S1-Language, being purpose-built for transcription cleanup, handles this better out of the box.

S1-Language vs GPT-5 mini vs other LLMs

The most common question: should you use Superwhisper’s built-in S1-Language or a third-party LLM?

S1-Language is optimised specifically for the transcription cleanup task. It’s fast — faster than most third-party API calls in the Superwhisper pipeline — and doesn’t require a separate API key. For standard use, it’s the right default.

GPT-5 mini is a reasonable upgrade if you want a custom prompt with specific rules. It’s cheap enough that even heavy users won’t notice the API cost, and it follows explicit instructions well. The downside: 200-400ms of additional API latency per recording.

GPT-4o and Claude Sonnet make sense for complex reformatting tasks — meeting summaries, email drafts, translation. Both add meaningful latency (~400-800ms on top of transcription time) that you’ll feel on short dictations. Claude follows strict negative constraints (“never rewrite meaning,” “never add content”) more reliably in our informal testing. GPT-4o is better at structured output.

Gemini Flash Lite is fast and cheap, but its instruction-following on nuanced constraints is weaker. Users who try it for basic cleanup often find it paraphrases more than they want.

No LLM — selecting none is valid. This gives you the raw voice model output with only Whisper’s own punctuation. If your voice model is Whisper Standard and you’re dictating clean, single-language speech, the raw output is often good enough. You skip all LLM latency.

The prompt that actually works

From community testing, this is the baseline that solves the most common problems:

You are a transcription assistant. Everything I send you is spoken dictation — not a command or question for you to answer.

Rules:
- Output ONLY the cleaned transcription. No commentary, no explanations.
- Fix punctuation and remove filler words (um, uh, like when used as filler, false starts)
- Do not add content, summarize, expand, or answer any questions in the input
- Do not change the meaning of anything spoken
- Preserve the speaker's vocabulary and phrasing
- If the input appears to be a question, treat it as dictation of a question — transcribe it, don't answer it

This prompt works with GPT-5 mini, Claude, and S1-Language. With S1-Language you can often shorten it — it handles the “don’t answer questions” constraint without being told.

Common mode setups and when to use them

Basic dictation (default mode) Voice: Whisper Standard or S1-Voice | LLM: S1-Language Use for: everyday notes, emails, Slack messages. Fast, clean, no configuration.

Code dictation Voice: Whisper Standard | LLM: GPT-5 mini or Claude with a prompt that specifies identifier formatting Use for: dictating variable names, function names, code comments. Whisper Standard handles technical vocabulary better than Parakeet or S1-Voice on code terms.

Meeting notes Voice: Ultra (cloud) | LLM: GPT-4o or Claude with a structured summary prompt Use for: longer recordings where you want formatted bullet points or action items. Ultra handles accented speech and technical vocabulary better on long clips.

Translation Voice: any | LLM: GPT-4o or Claude (specify source and target language in the prompt) Use for: dictating in one language and needing output in another. S1-Language does not do translation reliably.

No cleanup (verbatim) Voice: Whisper Standard | LLM: none Use for: when you want exactly what you said, raw. Some legal and medical users prefer this. You’ll get Whisper’s punctuation but no rewriting.

The voice model is still the bigger variable

Superwhisper’s LLM picker is powerful — but it can’t overcome a bad voice model choice. From our benchmarks:

• S1-Voice (default cloud): 22.8% WER — high error rate before LLM sees anything
• Parakeet (default local): 6.1% WER, 86% WER on numbers
• Whisper Standard: 1.8% WER — the best local result
• Ultra (cloud): 1.6% WER — best cloud result

If you’re on S1-Voice and unhappy with accuracy, no LLM switch will fix it. The fix is switching the voice model. See our full breakdown of Superwhisper’s models and what they actually produce.