Speech recognition guides
How to choose, run and measure speech-to-text models, with dated sources. Every number on these pages is measured or quoted from the vendor's own page.
Choosing a model
- The best open-source speech-to-text model for your use
Open speech models ranked on the seven Open ASR Leaderboard English sets, with size and languages. Which is best for accuracy, small size, or many languages.
- The best way to run speech-to-text on Apple silicon
MLX, Core ML on the Neural Engine, whisper.cpp, WhisperKit, FluidAudio and Apple's SpeechAnalyzer compared: which engine and model to use on an M-series Mac.
- Whisper alternatives: open models, faster runtimes and APIs
Alternatives to OpenAI Whisper by need: lower English error, smaller downloads, faster Whisper runtimes or hosted APIs. Measured numbers, and when Whisper wins.
- On-device speech-to-text for iOS, macOS, Android and desktop apps
How to add on-device speech recognition to an iOS, macOS, Android or desktop app: model sizes, runtimes such as Core ML, languages and the privacy trade-offs.
- Custom vocabulary in speech recognition: hotwords, keyterms and prompts
How custom vocabulary works in speech-to-text: hotwords, keyterms and prompts, with syntax and limits for Phonon, Deepgram, AssemblyAI, OpenAI and Whisper.
- How much does a speech-to-text API cost per hour?
Dated list prices per audio hour for speech-to-text APIs, from about $0.01 for hosted Whisper turbo to $0.96, and what self-hosting an open model costs.
- Self-hosted vs API speech-to-text: how to decide
When to run an open speech-to-text model yourself and when to pay a hosted API. Cost at your volume, latency, privacy, languages and features, side by side.
- Multilingual speech recognition for European languages
Word error by language on FLEURS for 25 European languages, from Spanish to Maltese, for three open models, and which open models cover which languages.
- Offline speech recognition in Python
Transcribe audio in Python without a cloud API: fermion (Phonon), openai-whisper, faster-whisper, Vosk and Moonshine compared, with install and run commands.
Model comparisons
- Phonon-2.5 vs Whisper
Phonon-2.5 is a 156 MB open speech model for 25 languages. How it compares with OpenAI Whisper on FLEURS word error, download size, languages and where it runs.
- Phonon-2.5 vs Parakeet TDT 0.6B v3 and Parakeet Redux
Per-language FLEURS word error for Phonon-2.5 (156 MB), Parakeet Redux (178 MB) and Parakeet TDT 0.6B v3 (2.5 GB), all scored with one code on 25 languages.
- Open speech-to-text models under 200 MB, compared
Five open speech recognition models between 156 and 178 MB compared on the seven Open ASR Leaderboard English sets, plus FLEURS for the two that cover 25 languages. Measured, dated.
- Self-host a speech-to-text server with an OpenAI-compatible endpoint
Run a 156 MB open speech model as a transcription server on CPU, Apple silicon or NVIDIA GPU. One command exposes POST /v1/audio/transcriptions; Docker images for CPU and CUDA.
- Speech recognition on the Apple Neural Engine with Core ML
Phonon-2 Core ML runs a 5.21 % word error open speech model on the Apple Neural Engine at 606 times realtime on an M5 MacBook Air, with Swift and Python runners. How to run it.
- The Open ASR Leaderboard and the Phonon models
What the Hugging Face Open ASR Leaderboard measures, how word error is normalised, and the Phonon rows: Phonon-2.5 and Phonon-2.5 English listed since November 2026.
- Phonon-2.5 vs Whisper large-v3-turbo
Phonon-2.5 averages 5.06 % English word error on seven Open ASR sets, Whisper large-v3-turbo 6.58 %, at 156 MB against 1,618 MB. Turbo covers 99 languages.
- Phonon-2.5 vs Deepgram Nova-3
Deepgram Nova-3 is a hosted speech API from $0.0043 a minute with diarization; Phonon-2.5 is a 156 MB open model you run yourself. Prices, features, trade-offs.
- Phonon-2.5 vs AssemblyAI Universal-3.5 Pro
AssemblyAI Universal-3.5 Pro has lower English word error than Phonon-2.5 (4.46 % vs 5.06 %) at $0.21 an hour; Phonon-2.5 is 156 MB and runs on your hardware.
- Phonon-2.5 vs ElevenLabs Scribe v2
ElevenLabs Scribe v2 beats Phonon-2.5 on English word error on all seven Open ASR sets (4.06 % vs 5.06 %). Phonon-2.5 runs on your hardware with no hourly fee.
- Phonon-2.5 vs Microsoft MAI-Transcribe-2
MAI-Transcribe-2 is Microsoft's hosted speech model on Azure: 60 languages, diarization, $0.10 an hour promo. Phonon-2.5 is a 156 MB open model you self-host.
- Phonon-2.5 vs Meta Muse Voice Transcribe
Muse Voice Transcribe: 4.93 % English word error; Phonon-2.5: 5.06 %; Phonon-2.5 English: 4.82 %. Muse is a $0.18 an hour API; Phonon runs on your hardware.
- Phonon-2.5 vs Mistral Voxtral
Phonon-2.5 averages 5.06 % English word error on seven Open ASR sets; open Voxtral models 5.21 to 6.12 %. Voxtral also answers questions about audio.
- Phonon-2.5 vs Moonshine
Moonshine's streaming models run from 34 M to 245 M parameters on phones, browsers and a Raspberry Pi. Phonon-2.5 is 156 MB with 25 European languages.
- Phonon-2.5 vs Qwen3-ASR
Qwen3-ASR 1.7B has lower English word error than Phonon-2.5 (4.43 % vs 5.06 %); Phonon-2.5 beats Qwen3-ASR 0.6B and covers 9 European languages Qwen does not.
- Phonon-2.5 vs NVIDIA Canary
Canary-1B-v2 covers Phonon-2.5's 25 languages and translates; Phonon-2.5 has lower English word error (5.06 % vs 5.81 %). Canary-Qwen 2.5B scores 4.55 %.
Benchmarks and methodology
- How Fermion Research measures speech models
Full public test sets, one scoring code for every model in a table, the leaderboard's English normaliser, diacritics kept for FLEURS, and our rows marked.
- FLEURS results in 25 languages
Phonon-2.5 averages 11.71 % word error on FLEURS across 25 languages, against 12.18 % for Parakeet Redux and 13.26 % for Parakeet TDT 0.6B v3. Every language.
- English speech recognition benchmark results
Phonon-2.5 English averages 4.82 % word error on the seven Open ASR Leaderboard English sets and 5.60 % on the current nine, scored with its code.
Deployment
- Replace expensive API model calls with a small model built for the task
Most production LLM calls do one narrow job. A small model trained for that job runs on your own hardware or on the device, answers in milliseconds, and cuts the per-call cost. How Fermion builds them.
- On-device and on-premises speech recognition for regulated teams
Transcribe and dictate without sending audio to a cloud API. Phonon runs on CPUs, Apple silicon and NVIDIA GPUs inside your environment; Detta dictates on the Mac's Neural Engine. Open weights, measured accuracy.