Best Local Speech-to-Text Models: 4 Tested on One File

Short answer: for English, NVIDIA Parakeet TDT 0.6B v3 wins — 6.32% average WER at 3,332x real-time, from a 600M-parameter model. Whisper large-v3 scores 7.44% at 145x real-time and wants ~10GB of VRAM for 1.55B parameters. Moonshine base is the CPU/streaming pick at 61.5M parameters. Voxtral Mini 3B is the one that also answers questions about the audio. And every single one of them roughly quintuples its error rate on accented speech.

Those WER and speed figures are not ours. They come from the Hugging Face Open ASR Leaderboard, which is the one place these models genuinely are run over the same audio — eight identical English test sets, one harness, every model. We pulled the raw result files from the leaderboard repository on 18 August 2026 rather than retyping numbers from a screenshot. Where a figure is a vendor's own claim, we say so.

What follows is the table, then the parts the table hides: which model falls apart on meetings, which one falls apart on dialect, which one cannot transcribe your language at all, and which one has a licence that will bite you if your company earns over a million dollars a year.

---

The Short Answer, By Situation

Pick by constraint, not by leaderboard position. Four of these models win outright in different rooms.

If you are choosing between just the top two, we have a dedicated Parakeet vs Whisper head-to-head that goes deeper on the architecture difference. This page is the category decision.

---

Same Audio, Every Model

On eight identical English test sets, Parakeet TDT 0.6B v3 leads on average WER and every Whisper variant trails it — while Voxtral Small 24B, at forty times Parakeet's size, only gets within 0.3 points.

Word Error Rate, lower is better. All figures from the Open ASR Leaderboard English short-form benchmark, retrieved 2026-08-18:

Source: huggingface/open_asr_leaderboard, scripts/data/en_shortform.csv, retrieved 2026-08-18. LS = LibriSpeech. *The leaderboard CSV rounds Whisper large-v3 to 2B; the openai/whisper README gives 1,550M, which is the figure we use elsewhere on this page.

Three things in that table are worth more than the ranking itself.

Parameter count has stopped predicting accuracy. A 600M model beats a 24B model on average. Moonshine streaming-medium, at 245M parameters, beats Whisper large-v3 at 1.55B. If you have been sizing your STT model the way you size an LLM, stop — the scaling story in ASR broke some time ago.

LibriSpeech Clean is a dead metric. Every model above 1B parameters lands between 1.59% and 2.10% on it. Clean, read, studio-recorded speech is solved. If a comparison page ranks models on LibriSpeech Clean, it is ranking noise. The columns that still separate models are AMI and Earnings22 — meetings and real business audio.

AMI is where the spread lives. From 10.68% (Moonshine streaming-medium) to 22.84% (Moonshine tiny) — a 12-point range on the same audio. AMI is multi-speaker meeting recordings with overlapping speech and far-field microphones, which is exactly the workload most people are actually buying an STT model for. Note the inversion: Voxtral Mini 3B, which is a strong model overall, is one of the worst here at 16.30%, behind a 245M-parameter Moonshine. Meeting audio is its own skill.

---

Speed: The 23x Gap

Parakeet transcribes at 3,332x real-time on the leaderboard harness; Whisper large-v3 manages 145x. That is not a tuning difference, it is an architecture difference.

Source: same leaderboard CSV, English short-form, 2026-08-18.

Whisper is an encoder-decoder transformer that generates text token by token, in 30-second windows. Parakeet is a FastConformer with a Token-and-Duration Transducer head, which predicts token durations and skips ahead — it does not pay a per-token autoregressive cost the way Whisper does. That is the whole gap.

The honest caveat on these numbers: RTFx here is measured with batching, on the leaderboard's evaluation hardware, over prepared datasets. You will not see 3,332x on your laptop transcribing one podcast episode, because a single file cannot fill a batch and your disk and audio decoding become the bottleneck. What survives to your machine is the ordering. Parakeet really is an order of magnitude faster than Whisper on the same GPU, and if you have a 40-hour backlog, that is the difference between an afternoon and a weekend.

For Whisper specifically there is a shortcut that closes part of the gap without changing models. SYSTRAN's own large-v2 benchmark for faster-whisper — a CTranslate2 reimplementation of the same weights — puts the reference openai/whisper at 2m23s and 4,708MB VRAM on their 13-minute test file, versus faster-whisper at int8 with batch_size=8 at 16s and 4,500MB, on an RTX 3070 Ti 8GB with CUDA 12.4. That is a vendor benchmark on a vendor test file, so treat the exact ratio with suspicion, but the direction is well replicated: if you are running stock openai/whisper you are leaving a lot on the table.

One warning on faster-whisper. The SYSTRAN repository was last pushed on 2025-11-19 — roughly nine months before this page went up — while NVIDIA-NeMo/Speech and moonshine-ai/moonshine were both pushed the same day we checked. It still works and it is still the pragmatic way to run Whisper, but it is no longer keeping pace with new models.

---

Where Each One Breaks

Every model on this page roughly quintuples its error rate on accented speech, and the leaderboard's long-form benchmark is the only place that shows it.

This is the section the listicles skip, because it requires looking past the headline average.

Accents and dialect: the five-fold penalty

The Open ASR Leaderboard's long-form benchmark includes CORAAL, a corpus of African American English. Compare each model's CORAAL score against its TEDLIUM score — same models, same benchmark run, wildly different audio:

Source: huggingface/open_asr_leaderboard, scripts/data/en_longform.csv, retrieved 2026-08-18.

Roughly one word in six is wrong on CORAAL for the best model here. If your audio is interviews, oral history, community recordings, customer calls, or anything that is not broadcast-register English, assume the marketing WER does not apply to you and budget for human correction. Parakeet is the least bad of the three, which is a genuine reason to prefer it beyond raw speed — but "least bad" is doing a lot of work in that sentence.

Notice also that the leaderboard publishes a separate "average without CORAAL" column, and on that metric the ordering flips: Whisper large-v3 leads at 8.67 against Parakeet's 9.10. Parakeet wins the long-form benchmark because it degrades less on dialect. Which model is "better" genuinely depends on whose voice you are transcribing.

Meetings and overlapping speakers

Covered in the table above, but to state it plainly: AMI errors run 4-6x higher than clean read speech for every model. None of these models do speaker diarisation on their own either — they will happily merge two people into one run-on paragraph. If you need "who said what", you need a diarisation layer on top; WhisperX is the usual answer, and our local meeting transcription walkthrough covers the full pipeline.

Long silences and hallucination

Whisper's well-known failure is inventing text during silence or music — it was trained on 30-second segments and will confidently fill an empty one, often with a subtitle-farm credit line. This is a Whisper-family architectural quirk, not a general ASR problem: transducer models like Parakeet emit nothing when there is nothing, because they have no reason to generate. If you are hitting repeated or invented phrases in Whisper output, that is a specific, fixable class of bug rather than a bad model.

Long files

Whisper chunks at 30 seconds and stitches, which is robust but is where timestamp drift creeps in. Parakeet TDT 0.6B v3's card states it handles up to 24 minutes with full attention (measured on an A100 80GB) or up to 3 hours with local attention — a vendor figure, and note the hardware it was measured on, since full-attention limits scale with available memory. Voxtral Mini's 32k context covers 30 minutes for transcription or 40 minutes for understanding, per Mistral's model card. For anything longer, all of them need a chunking strategy, and that is where your subtitle timings go wrong if you are careless.

---

Non-English

Parakeet ties Whisper large-v3 on the leaderboard's multilingual benchmark — but only across 25 European languages. Outside that set it produces nothing usable.

Source: scripts/data/multilingual.csv, Open ASR Leaderboard, retrieved 2026-08-18. Benchmark covers German, French, Italian, Spanish and Portuguese across CoVoST, FLEURS and MLS.

Read that table carefully, because the averages are misleading on their own. Parakeet matches Whisper large-v3's accuracy at 15x the speed — on the five languages this benchmark covers, all of which Parakeet supports. Its full supported set is Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Russian and Ukrainian. It auto-detects the language without prompting. If your audio is Japanese, Arabic, Hindi, Mandarin or any of the seventy-odd other languages Whisper covers, Parakeet is not a candidate at all and this benchmark tells you nothing.

Two more constraints worth knowing. Whisper large-v3-turbo cannot translate — the openai/whisper README states it "is not trained for translation tasks" and will return the original language even if you pass --task translate. Use large or medium for that. And Voxtral's multilingual strength is concentrated in the eight languages Mistral names (English, Spanish, French, Portuguese, Hindi, German, Dutch, Italian); it is a different shape of model, and our Voxtral local guide covers what its audio-understanding mode adds beyond plain transcription.

---

The Licensing Trap

Moonshine's README says MIT. Its LICENSE file says the non-English models are non-commercial above $1M in revenue. Read the file, not the README.

This is the one thing on this page that can actually cost you money, and it is why GitHub reports the moonshine-ai/moonshine repository licence as NOASSERTION rather than MIT.

The LICENSE file opens with an explicit split:

"The code in this repo ... is licensed under the MIT License. The English-language models are also released under the MIT License. See SECTION 1 for terms. Models for other languages are released under the Moonshine Community License, which is a non-commercial license. See SECTION 2 for terms."

Section 2 is the Moonshine AI Community License Agreement (last updated 15 June 2025), which grants free research and non-commercial use, and preserves free commercial access only for entities "generating annual revenue of less than US $1,000,000 (or local currency equivalent)". Above that threshold you need to talk to them.

Here is the full picture:

Licences read from each repository's LICENSE file and Hugging Face model card metadata on 2026-08-18.

Nothing about this is sneaky — Moonshine documents it in the file — but "MIT" in a README is what most people check, and it is wrong for two-thirds of what they ship. If you are building a commercial product on non-English Moonshine weights, that is the sentence to take to whoever signs off on your dependencies.

---

Pick One by VRAM

Below 6GB you have better options than Whisper. Above 10GB the choice is about language coverage, not capacity.

The openai/whisper README's own size table is the cleanest reference point anyone publishes for the Whisper family: tiny 39M (~1GB), base 74M (~1GB), small 244M (~2GB), medium 769M (~5GB), large 1,550M (~10GB), turbo 809M (~6GB), with turbo described as "an optimized version of large-v3 that offers faster transcription speed with a minimal degradation in accuracy". The leaderboard puts a number on "minimal": 7.83% versus 7.44% average WER, for 4GB less VRAM. That is a good trade for most people, as long as you do not need translation.

If you are still deciding what hardware to run any of this on, our Whisper local setup guide has the practical install path, and the best local TTS models roundup covers the other direction if you are building a full voice pipeline.

---

Verified Install Commands

Every command below is copied from the project's own README or model card, not reconstructed from memory.

Whisper (MIT, ~99 languages):

pip install -U openai-whisper
whisper audio.mp3 --model turbo

faster-whisper (same weights, less VRAM):

pip install faster-whisper

from faster_whisper import WhisperModel
model = WhisperModel("large-v3", device="cuda", compute_type="int8_float16")
segments, info = model.transcribe("audio.mp3")

Parakeet TDT 0.6B v3 — GPU path via NeMo:

pip install -U nemo_toolkit['asr']

import nemo.collections.asr as nemo_asr
asr_model = nemo_asr.models.ASRModel.from_pretrained(model_name="nvidia/parakeet-tdt-0.6b-v3")
output = asr_model.transcribe(['audio.wav'])
print(output[0].text)

Parakeet on CPU — the GGUF path, via NVIDIA's native C++ runtime:

hf download nvidia/parakeet-tdt-0.6b-v3 \
  parakeet-tdt-0.6b-v3.q8_0.gguf \
  --local-dir models

nemo-speech transcribe audio.wav \
  --model models/parakeet-tdt-0.6b-v3.q8_0.gguf

Moonshine (English MIT; see licensing above for other languages):

pip install moonshine-voice
moonshine-voice mic --language en

Voxtral Mini 3B (Apache-2.0, needs ~9.5GB):

uv pip install -U "vllm[audio]" --system
vllm serve mistralai/Voxtral-Mini-3B-2507 \
  --tokenizer_mode mistral --config_format mistral --load_format mistral

Parakeet emits automatic punctuation and capitalisation plus word-, segment- and character-level timestamps out of the box — pass timestamps=True to transcribe(). That matters more than it sounds: it removes the separate alignment step that Whisper pipelines usually bolt on.

---

What This Page Does Not Prove

We did not run these models ourselves. Every WER and RTFx number here is from the Hugging Face Open ASR Leaderboard, which is a third-party harness running all of these models over the same eight English test sets — that is what makes "the same audio" a defensible claim rather than a slogan. But it is their hardware, their batching, their preprocessing. We pulled the raw CSVs rather than transcribing a screenshot, and we have linked them so you can check every cell.

Three specific limits worth stating:

The leaderboard is English-first. Its multilingual benchmark covers five languages. Whisper's ~99-language coverage and Voxtral's eight are largely untested by any public same-audio comparison, so for languages outside German, French, Italian, Spanish and Portuguese, you are back to testing on your own audio.

Nothing here measures your recording conditions. The gap between AMI (11-23% WER) and LibriSpeech Clean (1.6-4.6%) on identical models is larger than the gap between any two models on this page. Your microphone, room and speakers will move your results more than your model choice will. Transcribe ten minutes of your actual audio with two candidates before you commit to either.

faster-whisper's numbers are the vendor's own. The 2m23s-to-16s comparison comes from SYSTRAN's README, on their file, on an RTX 3070 Ti. We reproduce it as an attributed vendor claim because there is no neutral equivalent, not as an independent measurement.

Recheck date: mid-November 2026. NVIDIA-NeMo/Speech and moonshine-ai/moonshine were both pushed to on the same day we ran these checks (2026-08-18), and the leaderboard adds models continuously. The ordering here has been stable for a while, but the leaderboard CSVs linked below are always more current than this page.

---

Sources

• huggingface/open_asr_leaderboard — scripts/data/en_shortform.csv, en_longform.csv and multilingual.csv, retrieved 2026-08-18; all WER, RTFx and CORAAL figures
• openai/whisper README — model size / VRAM / relative speed table, turbo translation limitation (MIT, ~107,500 stars)
• nvidia/parakeet-tdt-0.6b-v3 model card — 600M params, 25 supported languages, CC BY 4.0, 24-minute full-attention limit, 2GB RAM floor, GGUF and NeMo usage
• moonshine-ai/moonshine LICENSE — Section 1 (MIT, English models) and Section 2 (Moonshine AI Community License, $1M revenue threshold)
• mistralai/Voxtral-Mini-3B-2507 model card — ~9.5GB bf16/fp16, 32k context, 30-minute transcription window, Apache-2.0
• SYSTRAN/faster-whisper README — CTranslate2 benchmark on RTX 3070 Ti 8GB and Intel i7-12700K (vendor benchmark)
• Repository metadata via the GitHub API, 2026-08-18: openai/whisper ~107,500 stars; SYSTRAN/faster-whisper ~25,000 (last push 2025-11-19); NVIDIA-NeMo/Speech ~18,200 (last push 2026-08-18); moonshine-ai/moonshine ~10,900 (last push 2026-08-18). Star counts drift daily and are given as approximations.

---

FAQ