llms.txt
# Codictate
> Codictate is a free, open-source dictation app for macOS (Apple Silicon) and Windows that turns speech into text entirely on your device. Press a shortcut, speak, and the text is pasted wherever your cursor already is: an IDE, a browser, a chat window, notes, or email. The differentiator is that nothing leaves the machine. There is no account, no cloud transcription, and no telemetry, because the speech models (Whisper, Parakeet TDT v3, and the Danish models Edda v0.2 and hviske-v5-tiny) run locally.
A typical session never interrupts what you were doing. You keep the caret where you want the words, hold the dictation shortcut, talk, and release; Codictate transcribes on-device and pastes the result into the focused field. A user dictionary rewrites the terms local models routinely mishear, so a spoken name or address lands correctly instead of needing a manual fix. Models are swappable from Settings, which is the practical trade-off this product asks you to make: Parakeet TDT v3 for speed on clean English, a Whisper Large V3 variant when you switch languages or work in noise, Edda v0.2 for the most accurate Danish, and a small quantized model when memory is tight. Every one of those claims is measured rather than asserted; the benchmark runs below publish the raw accuracy, speed, and memory numbers. Codictate is aimed at developers and other keyboard-heavy desktop users who want dictation without a subscription and without sending their voice to a server. It is licensed under Apache 2.0, and an iOS companion app is on the App Store.
## Primary
- [Codictate](https://codictate.app/): Home page: what Codictate does, the demo, and the download for macOS and Windows
- [Codictate for iPhone](https://codictate.app/ios): The iOS app: a keyboard Dictate button with Parakeet TDT v3 running on the phone, free on the App Store
- [About](https://codictate.app/about): Why Codictate exists: who builds it, what on-device means in practice, and how it differs from a cloud dictation tool
- [Releases](https://codictate.app/releases): Every published release with its notes and download assets, pulled from GitHub
- [Support](https://codictate.app/support): Report a bug or request a feature; the form files an issue on the GitHub tracker, no account needed
## Documentation
- [Blog](https://codictate.app/blog): Guides on dictating from a Mac or Windows PC, model deep dives, and release notes in long form
- [Edda v0.2 is now Codictate's most accurate Danish model](https://codictate.app/blog/edda-v0-2-danish-model): Edda v0.2, a Danish Whisper model from the Alexandra Institute, is in Codictate. On FLEURS Danish it got 7.53% of words wrong vs 11.31% for hviske.
- [Wispr shipped a new speech model the day I finished testing the old one. So I ran all 8,287 recordings again.](https://codictate.app/blog/wispr-flow-canto-rerun): Wispr Flow switched to Canto on 17 September 2026, so I reran all 8,287 recordings. Accuracy barely moved. Flow got 17 to 26% faster in every language.
- [We ran Wispr Flow on every recording we have. 8,287 of them, 19.8 hours of audio.](https://codictate.app/blog/wispr-flow-full-test-suite): Wispr Flow 1.6.897 on all 8,287 recordings in our test set: 96.00% in clean English, 77.94% in Hungarian. Codictate's local models win 4 of 5 languages.
- [We measured Wispr Flow in five of its 100+ languages. English scored 95.5%, Hungarian 76.7%.](https://codictate.app/blog/wispr-flow-accuracy-benchmark): We played 2,000 LibriSpeech and FLEURS recordings into Wispr Flow 1.6.774. It pooled 91.76% word accuracy and beat every model we ship on Danish.
- [A 153 MB Danish model beats Whisper Large V3 at Danish dictation](https://codictate.app/blog/danish-dictation-hviske-v5-tiny): Hviske v5 Tiny is now in Codictate. On FLEURS Danish it got 88.7% of words right vs 84.8% for Whisper Large V3 Turbo, 3x faster on a third of the memory.
- [We swapped whisper-cli for CrispASR: 1.58x faster, 24% less RAM, and 107 backends instead of one](https://codictate.app/blog/crispasr-replaces-whisper-cli): CrispASR, the whisper.cpp fork with 100+ backends, vs whisper-cli on an M4 Max: 1.58x faster, 24% less RAM, same accuracy. What it is and when to switch.
- [Tiny vs Base: a 50-sample triage to find which Whisper models deserve a full benchmark](https://codictate.app/blog/tiny-vs-base-triage): 12 Whisper Tiny and Base variants on an M4 Max: English vs multilingual accuracy, RAM and speed. Which small model is worth using.
- [We benchmarked every model in Codictate. Here is what 4,000 transcriptions revealed.](https://codictate.app/blog/stt-model-benchmarks): Whisper Small, Large V3, Large V3 Turbo and Parakeet TDT v3 on an M4 Max: accuracy in 5 languages, speed and RAM. Parakeet is 8x faster than Large V3.
- [Wispr Flow Alternative for MacBook and Windows](https://codictate.app/blog/wispr-flow-alternative-macbook): A free, open-source Wispr Flow alternative for Mac and Windows. Runs offline, no subscription, and beat Flow in 4 of 5 languages on 8,287 recordings.
- [Free Dictation Software for Mac (Works in Any App)](https://codictate.app/blog/free-dictation-tool-mac): Free dictation software for Mac and MacBook: open source, offline, no word limits. Dictate into any app with Option+Space. Compared with Apple Dictation.
- [How to Use Dictation on Your Mac](https://codictate.app/blog/how-to-use-dictation-on-your-mac): Turn on Mac dictation, change the shortcut, pick languages, and fix it when it stops working. Plus a free offline app for longer text and code editors.
- [NVIDIA Parakeet TDT v3: dictation that stays on your Mac and iPhone](https://codictate.app/blog/nvidia-parakeet-tdt-v3): Run NVIDIA Parakeet TDT 0.6B v3 offline on Mac (Core ML, Neural Engine) and iPhone. 25 languages, 11 ms per second of audio, 80 MB RAM. Free.
- [Translate Mode + two days after launch](https://codictate.app/blog/translate-mode-launch-update): Thank you for the launch week energy we hit #27 on Product Hunt. New: speak in your language, get English text. Plus notes on Apple code signing.
- [Any language. Any application. Anytime.](https://codictate.app/blog/any-language-any-application-anytime): Why I built Codictate: Electrobun, a Wispr Flow trial ending, and a rule: no AI until I could dictate my prompts with my own tool.
- [Benchmarks](https://codictate.app/benchmarks): The head to head against Wispr Flow per language, accuracy and speed for every model Codictate can run, which model to pick, and how the runs are measured
- [How we measure](https://codictate.app/benchmarks/methodology): Method behind every benchmark figure: the clip set, the fingerprint that proves both products got the same clips, pooled Word Error Rate, how speed is timed on each side, and how a cloud product with no API was driven through its own app
- [Speech models](https://codictate.app/models): Which speech model to pick for dictation on a Mac, with accuracy, speed, RAM and file size for each
- [Model guide: Whisper Large V3 Turbo Q5_0](https://codictate.app/models/whisper-large-v3-turbo-q5-0): OpenAI's Whisper Large V3 Turbo, quantized to 5 bits in the whisper.cpp GGML format. It is Codictate's default model and ships with the app.
- [Model guide: Whisper Large V3 Q5_0](https://codictate.app/models/whisper-large-v3-q5-0): OpenAI's Whisper Large V3, quantized to 5 bits in the whisper.cpp GGML format. Unlike the Turbo builds it can translate into English.
- [Model guide: Whisper Large V3](https://codictate.app/models/whisper-large-v3): OpenAI's Whisper Large V3 at full precision in the whisper.cpp GGML format. The reference build the quantized versions are compared against.
- [Model guide: Whisper Large V3 Turbo](https://codictate.app/models/whisper-large-v3-turbo): OpenAI's Whisper Large V3 Turbo at full precision in the whisper.cpp GGML format: Large V3 with a much smaller decoder.
- [Model guide: Whisper Medium Q5_0](https://codictate.app/models/whisper-medium-q5-0): OpenAI's multilingual Whisper Medium, quantized to 5 bits in the whisper.cpp GGML format.
- [Model guide: Whisper Small Q5_1](https://codictate.app/models/whisper-small-q5-1): OpenAI's multilingual Whisper Small, quantized in the whisper.cpp GGML format. The small option for a Mac that is short on memory.
- [Model guide: Parakeet TDT 0.6B v3](https://codictate.app/models/parakeet-tdt-0-6b-v3): NVIDIA's Parakeet TDT 0.6B v3, a speech model for 25 European languages. On a Mac, Codictate runs a Core ML conversion on the Apple Neural Engine. It is also the default engine in Codictate for iPhone.
- [Model guide: Hviske v5 Tiny Q5_0](https://codictate.app/models/hviske-v5-tiny-q5-0): syvai's hviske v5 tiny, a Danish-only speech model, in its 5-bit GGUF build. It runs under CrispASR's cohere backend. The model is licensed CC BY-NC 4.0, non-commercial.
- [Model guide: Edda v0.2 Q5_0](https://codictate.app/models/edda-v0-2-q5-0): Edda v0.2 is Whisper Large V3 Turbo trained further on Danish by the Alexandra Institute, released under Apache 2.0. Codictate converted it to the whisper.cpp GGML format and quantized it to 5 bits. Danish only.
- [Benchmark run: Edda v0.2 on Danish: F16, Q8_0 and Q5_0](https://codictate.app/benchmarks/edda-v0-2-danish): Edda v0.2, a Danish fine-tune of Whisper Large V3 Turbo, in F16, Q8_0 and Q5_0 on all 927 FLEURS Danish recordings: accuracy, speed and RAM on an M4 Max.
- [Benchmark run: Tail run: English recordings 951 to the end, 13 models](https://codictate.app/benchmarks/2026-09-v4-english-tail): Partial run that takes 13 Whisper and Parakeet models through the rest of LibriSpeech. The pooled results are on the benchmarks page. (Partial run.)
- [Benchmark run: Wispr Flow 1.6.897 English tail](https://codictate.app/benchmarks/2026-09-v5-english-tail): Partial run: the rest of LibriSpeech on Wispr Flow 1.6.897. Pooled with the five-language run on the benchmarks page. (Partial run.)
- [Benchmark run: Wispr Flow 1.6.897 (Canto): all five languages](https://codictate.app/benchmarks/2026-09-v5-full-1-6-897): Wispr Flow 1.6.897, the first version on Canto, on English, Spanish, Danish and Hungarian. Pooled with the English tail run on the benchmarks page. (Partial run.)
- [Benchmark run: Catch-up run: Danish recordings 401 to 927, Hviske v5 Tiny](https://codictate.app/benchmarks/2026-09-v4-hviske-catchup): Partial run that extends the five Hviske v5 Tiny quantizations to the end of FLEURS Danish. The pooled results are on the benchmarks page. (Partial run.)
- [Benchmark run: Catch-up run: recordings 401 to 950, 13 models](https://codictate.app/benchmarks/2026-09-v4-multilingual-catchup): Partial run that extends 13 Whisper and Parakeet models from 400 to 950 recordings per language. The pooled results are on the benchmarks page. (Partial run.)
- [Benchmark run: Wispr Flow English: recordings 951 to 2,200](https://codictate.app/benchmarks/2026-09-v4-english): Partial run: LibriSpeech recordings 951 to 2,200 played into Wispr Flow. The pooled Wispr Flow results are on the benchmarks page. (Partial run.)
- [Benchmark run: Wispr Flow Danish recheck: recordings 401 to 927](https://codictate.app/benchmarks/da-recheck-1-6-827): Partial run: the second half of FLEURS Danish played into Wispr Flow again. The pooled Wispr Flow results are on the benchmarks page. (Partial run.)
- [Benchmark run: Hviske v5 Tiny Q4_K benchmark](https://codictate.app/benchmarks/2026-09-v2-hviske-v5-tiny-q4-k): Word accuracy and speed of Hviske v5 Tiny Q4_K in Codictate on an Apple M4 Max, 400 recordings per language: Danish.
- [Benchmark run: Hviske v5 Tiny Q5_0 benchmark](https://codictate.app/benchmarks/2026-09-v2-hviske-v5-tiny-q5-0): Word accuracy and speed of Hviske v5 Tiny Q5_0 in Codictate on an Apple M4 Max, 400 recordings per language: Danish.
- [Benchmark run: Hviske v5 Tiny Q6_K benchmark](https://codictate.app/benchmarks/2026-09-v2-hviske-v5-tiny-q6-k): Word accuracy and speed of Hviske v5 Tiny Q6_K in Codictate on an Apple M4 Max, 400 recordings per language: Danish.
- [Benchmark run: Hviske v5 Tiny Q8_0 benchmark](https://codictate.app/benchmarks/2026-09-v2-hviske-v5-tiny-q8-0): Word accuracy and speed of Hviske v5 Tiny Q8_0 in Codictate on an Apple M4 Max, 400 recordings per language: Danish.
- [Benchmark run: Hviske v5 Tiny F16 benchmark](https://codictate.app/benchmarks/2026-09-v2-hviske-v5-tiny-f16): Word accuracy and speed of Hviske v5 Tiny F16 in Codictate on an Apple M4 Max, 400 recordings per language: Danish.
- [Benchmark run: Whisper Medium Q8_0 benchmark](https://codictate.app/benchmarks/2026-09-v2-medium-q8-0): Word accuracy and speed of Whisper Medium Q8_0 in Codictate on an Apple M4 Max, 400 recordings per language: English, Spanish, Danish and Hungarian.
- [Benchmark run: Whisper Medium Q5_0 benchmark](https://codictate.app/benchmarks/2026-09-v2-medium-q5-0): Word accuracy and speed of Whisper Medium Q5_0 in Codictate on an Apple M4 Max, 400 recordings per language: English, Spanish, Danish and Hungarian.
- [Benchmark run: Whisper Medium benchmark](https://codictate.app/benchmarks/2026-09-v2-medium): Word accuracy and speed of Whisper Medium in Codictate on an Apple M4 Max, 400 recordings per language: English, Spanish, Danish and Hungarian.
- [Benchmark run: Whisper Large V2 Q8_0 benchmark](https://codictate.app/benchmarks/2026-09-v2-large-v2-q8-0): Word accuracy and speed of Whisper Large V2 Q8_0 in Codictate on an Apple M4 Max, 400 recordings per language: English, Spanish, Danish and Hungarian.
- [Benchmark run: Whisper Large V2 Q5_0 benchmark](https://codictate.app/benchmarks/2026-09-v2-large-v2-q5-0): Word accuracy and speed of Whisper Large V2 Q5_0 in Codictate on an Apple M4 Max, 400 recordings per language: English, Spanish, Danish and Hungarian.
- [Benchmark run: Whisper Large V2 benchmark](https://codictate.app/benchmarks/2026-09-v2-large-v2): Word accuracy and speed of Whisper Large V2 in Codictate on an Apple M4 Max, 400 recordings per language: English, Spanish, Danish and Hungarian.
- [Benchmark run: Whisper Large V1 benchmark](https://codictate.app/benchmarks/2026-09-v2-large-v1): Word accuracy and speed of Whisper Large V1 in Codictate on an Apple M4 Max, 400 recordings per language: English, Spanish, Danish and Hungarian.
- [Benchmark run: Parakeet TDT 0.6B v3 benchmark](https://codictate.app/benchmarks/2026-09-v2-parakeet-tdt-0-6b-v3): Word accuracy and speed of Parakeet TDT 0.6B v3 in Codictate on an Apple M4 Max, 400 recordings per language: English, Spanish, Danish and Hungarian.
- [Benchmark run: Whisper Large V3 Turbo Q8_0 benchmark](https://codictate.app/benchmarks/2026-09-v2-large-v3-turbo-q8-0): Word accuracy and speed of Whisper Large V3 Turbo Q8_0 in Codictate on an Apple M4 Max, 400 recordings per language: English, Spanish, Danish and Hungarian.
- [Benchmark run: Whisper Large V3 Turbo Q5_0 benchmark](https://codictate.app/benchmarks/2026-09-v2-large-v3-turbo-q5-0): Word accuracy and speed of Whisper Large V3 Turbo Q5_0 in Codictate on an Apple M4 Max, 400 recordings per language: English, Spanish, Danish and Hungarian.
- [Benchmark run: Whisper Large V3 Turbo benchmark](https://codictate.app/benchmarks/2026-09-v2-large-v3-turbo): Word accuracy and speed of Whisper Large V3 Turbo in Codictate on an Apple M4 Max, 400 recordings per language: English, Spanish, Danish and Hungarian.
- [Benchmark run: Whisper Large V3 benchmark](https://codictate.app/benchmarks/2026-09-v2-large-v3): Word accuracy and speed of Whisper Large V3 in Codictate on an Apple M4 Max, 400 recordings per language: English, Spanish, Danish and Hungarian.
- [Benchmark run: Whisper Large V3 Q5_0 benchmark](https://codictate.app/benchmarks/2026-09-v2-large-v3-q5-0): Word accuracy and speed of Whisper Large V3 Q5_0 in Codictate on an Apple M4 Max, 400 recordings per language: English, Spanish, Danish and Hungarian.
- [Benchmark run: Wispr Flow 1.6.774 on 400 recordings per language](https://codictate.app/benchmarks/2026-09-v2-flow): Wispr Flow 1.6.774, driven through its own Mac app, on the same 400 recordings per language as Codictate's models: English, Spanish, Danish and Hungarian.
- [Benchmark run: Hviske v5 Tiny Q5_0 on 400 Danish recordings](https://codictate.app/benchmarks/hviske-danish-400): Hviske v5 Tiny Q5_0 on 400 FLEURS Danish recordings, the depth of the first Wispr Flow comparison, in Codictate on an Apple M4 Max.
- [Benchmark run: Codictate's main models at 400 recordings per language](https://codictate.app/benchmarks/curated-400-wispr-comparison): Whisper Small, Large V3 Turbo, Large V3 and Parakeet TDT 0.6B v3 on 400 recordings per language, the depth of the first Wispr Flow comparison. M4 Max.
- [Benchmark run: Hviske v5 Tiny vs Whisper and Parakeet on Danish](https://codictate.app/benchmarks/hviske-vs-main-models): All five Hviske v5 Tiny quantizations against Whisper Large V3 Turbo Q5_0 and Parakeet TDT 0.6B v3 on about 200 FLEURS Danish recordings, on an M4 Max.
- [Benchmark run: CrispASR vs whisper-cli: the same Whisper models, two runtimes](https://codictate.app/benchmarks/crispasr-vs-whisper): Whisper Large V3, Large V3 Turbo and Medium English (Q5_0) under whisper-cli and under CrispASR on an M4 Max: speed, peak RAM and word accuracy.
- [Benchmark run: 34 Whisper and Parakeet models: the screening run](https://codictate.app/benchmarks/full-run-except-tiny-base): Every Whisper size from Tiny to Large V3 Turbo, quantized and full precision, plus Parakeet TDT 0.6B v3, on 50 recordings per language on an Apple M4 Max.
- [Benchmark run: Whisper Tiny and Base triage: 12 variants](https://codictate.app/benchmarks/tiny-base-triage): All 12 Whisper Tiny and Base variants (multilingual, English-only, Q5_1, Q8_0) on 50 recordings per language on an M4 Max, to see which deserve a full run.
- [Benchmark run: Whisper vs Parakeet: the first main-model comparison](https://codictate.app/benchmarks/main-model-comparison): Whisper Small, Large V3 Turbo and Large V3 (quantized) against Parakeet TDT 0.6B v3 in Codictate: accuracy, speed and RAM in 5 languages on an M4 Max.
## Legal
- [Privacy Policy](https://codictate.app/privacy): No accounts, no telemetry, no cloud processing: what the app does and does not touch
- [Terms of Service](https://codictate.app/terms): Apache 2.0 licence terms for using the app, provided as-is with no account required
## Optional
- [Sitemap](https://codictate.app/sitemap.xml): Machine-readable list of public pages
