Voice input

Turn speech into text ready to enter into an app. Compare with Typeless and Aqua Voice and check where each processes audio.

Selection guide

Models by task

Start with

Whisper large-v3-turbo

MLX / FP16

Starting memory estimate
16 GB

For trying short tasks one at a time

The official ~6GB figure describes GPU memory, not total Mac RAM. The 16GB starting estimate includes the OS and audio app. Split recordings and run the summarizer separately.

Tasks to try

Transcribe recordings. For meeting notes, pass the transcript to a text model such as Qwen3.5-9B afterward.

Keep under review

Speaker diarization is a separate feature. Check noise, overlapping speech and text invented during silence.

Setup and limits

Getting started · Budget rationale

Memory figures are editorial budgets including room for the OS and apps, not verified minimums. Text models assume 4-bit weights, 4K–8K input tokens and one model at a time. Long documents, images and concurrent apps need more. A3B does not mean the entire model fits in 3GB.

Assumes an Apple Silicon Mac and a compatible app: LM Studio or similar for text, Whisper apps for audio, and model-specific image/video tools. RAM alone does not determine speed. Start with a short input on your current Mac.

User feedback overview

What users value

Transcription speed is valued.

Recurring difficulties

Users report hallucinated text in silence and variable accuracy by language and recording. VAD can help but does not remove the need to review.

A qualitative synthesis of public discussions, not a representative survey. Reports include PCs and cloud setups, so they do not establish Mac speed or memory requirements. No individual profiles or anecdotes are reproduced.

Our tests and conditions

Candidates to test

On hold or being checked again Speech-to-text failed; tidying the text has passed, the full task remains unverified · Updated

On 2026-11-15, Qwen3.6 35B-A3B on a Mac Studio (M5 Ultra, 96GB) was tested once on tidying speech-recognition output into text ready to enter (consistent number style, keeping English words, not changing what was heard). It passed all 11 grading checks in about 8 seconds with no manual fixes. Qwen3.8 27B also passed the same kind of task on 11/14. Speech-to-text itself failed on 11/6 and 11/8 and has not been measured since. A full run from speech-to-text through tidying, and the amount of memory used, have not been checked yet.

Whisper large-v3-turbo · Qwen3-ASR 1.7B · Qwen3.5 4B / 9B · Gemma 4 12B

Test tasks and measures

Use the same Japanese, English and Chinese recordings. Compare recognition errors, punctuation and name corrections, waiting time and steps to enter the text.

What we know and what remains open

Recognition and rewriting are evaluated separately. Mac compatibility and the input workflow still need testing.

Before trying it on your Mac

On hold or being checked again

Run: 20261106T212629-mac-llm-bench-a06d · Input SHA-256: 7d3514b0d992c62531467aa46b9217bef83f061f6ace8c750658b4b824616002

Mac mini M6, 16 GB; mlx-whisper 0.4.3; mlx-community/whisper-large-v3-turbo (revision a4aaeec0636e6fef84abdcbe3544cb2bf7e9f6fb); Japanese transcription, temperature 0. Inference took 18.173797 s. Peak memory was not measured because the runner has no sampler.

One inference run. Input: voice-asr-synthetic-ja-v1 (Japanese). Transcript: “明日の会議は10時からです。資料は3部用意してください。メールはベンチマークのexample.comに送ってください。” CER: 0.105263. Independent grading failed because “十時” (ten o’clock) and “三部” (three copies) did not match the required phrases. No transcript editing was performed.

Response time: 18.1738 s · CER: 0.105263 · Peak memory: No measurements yet

The earlier run (queue 20261106T185152-mac-llm-bench-fcfa) stopped after 4.265656 s because ffmpeg was unavailable; it produced no transcript. This failed attempt remains in the record.

This fixed case records a failed speech-recognition run. End-to-end replacement, including number normalization or rewriting, has not been tested. The next cycle should use a task that checks recognition separately from post-processing, informed by the number and proper-name errors observed here.

Cycle 2 measurement (failed)

Run: 20261108T115454-mac-llm-bench-5254 · Input SHA-256: 7d3514b0d992c62531467aa46b9217bef83f061f6ace8c750658b4b824616002

Mac mini, 16 GB; mlx-whisper 0.4.3; mlx-community/whisper-large-v3-mlx (revision 49e6aa286ad60c14352c404340ded53710378a11); Japanese transcription, temperature 0. Input SHA-256: 7d3514b0d992c62531467aa46b9217bef83f061f6ace8c750658b4b824616002.

Queue 20261108T115454-mac-llm-bench-5254, 2026-11-08 11:54:58–11:55:13 JST. Transcript: “明日の会議は10時からです。資料は3部用意してください。メールはbenchmarkのexample.comに送ってください。” Adapter inference time was 8.675082 s. Independent grading failed (CER 0.2631578947, above threshold); required phrases “十時” and “三部” did not match. No edits were made. Peak memory was not measured.

The separate cycle 1 run (2026-11-06, whisper-large-v3-turbo, 18.173797 s) also failed independent grading (CER 0.105263; two required phrases mismatched). Its result remains in a separate record.

This ASR-only failure does not establish full voice-input replacement quality. The next cycle should choose a task that measures recognition separately from post-processing, informed by the numeral and proper-name errors.

Cycle 3 measurement (post-processing passed, full use case held)

Run: 20261114T155631-mac-llm-bench-61fc · Input SHA-256: 41f540063da27ec8b4b8b526e25594365aea2f2c8769ce7f509108288c8546da

Mac Studio No. 1; Qwen3.8 27B (Q4_K_M, digest aaee06c39dcf2437cde036998d960e1fc1494b8191be7cc9657d01e509097813); fixed case voice-postprocess-001. ollama-chat 0.34.3, temperature 0, seed 42, num_ctx 4096, num_predict 512. Input SHA-256: 41f540063da27ec8b4b8b526e25594365aea2f2c8769ce7f509108288c8546da.

Queue 20261114T155631-mac-llm-bench-61fc, 2026-11-14 15:56:34–15:56:37 JST. Inference time was 3.417818 s. Passed all 10 independent checks (kanji numeral normalization and English text preservation). No manual edits. Peak memory was not measured because the runner is uninstrumented. One case does not establish performance for the full use case.

Cycle 1 (2026-11-06, Whisper Turbo failed, 18.173797 s) and cycle 2 (2026-11-08, Whisper Large v3 MLX failed, 8.675082 s) remain recorded as separate attempts. Status remains held.

Next steps include integrated speech recognition and post-processing, multi-model and multilingual comparisons, app-entry workflow integration, and peak memory measurement.

Cycle 4 measurement (post-processing passed, full use case held)

Run: 20261115T161327-mac-llm-bench-d1d1 · Input SHA-256: 182ab18da6a4e4b21ce6f80d754a774d1616d5b934ea37d01bee5a1b28a90b6d

Mac Studio 2 (96 GB); Qwen3.6 35B-A3B (digest a7eb95c53bcf96b4bdd008d0fab4a5dac88047d9c1a7a9ab88ed453423fbd87c); fixed case voice-postprocess-mem-001. ollama-chat 0.40.5, temperature 0, seed 42, num_ctx 4096, num_predict 512, think false, format json. Input SHA-256: 182ab18da6a4e4b21ce6f80d754a774d1616d5b934ea37d01bee5a1b28a90b6d.

Queue 20261115T161327-mac-llm-bench-d1d1, 2026-11-15 16:13:32–16:13:42 JST. Inference time was 8.363363 s. All 11 grader checks passed (kanji numeral normalization, English text preservation, recognition text unchanged). No manual edits. The recorded peak of 46,039,040 bytes is the maximum RSS of the Ollama process only, which excludes model weights; it is not treated as effective memory and remains unconfirmed. One case does not establish performance for the full use case.

Cycle 3 (11/14, Qwen3.8 27B on Mac Studio 1, voice-postprocess-001, 3.418 s) remains a separate post-processing record. Because the model and machine differ, its speed and accuracy cannot be compared with this cycle's results. The speech-recognition failures from cycle 1 (11/6) and cycle 2 (11/8) also remain separate records. Status remains held.

Speech recognition (ASR) is still not included. Passing the full use case still requires an integrated ASR-and-postprocessing test, multiple fixed cases, a peak-memory method that includes model weights, and a check of the input-into-app steps.