# Whisper vs Parakeet > Whisper vs Parakeet in October 2026: OpenAI's multilingual ASR versus NVIDIA's fast transducer on accuracy, speed, word timing and hallucination. Source: https://metavert.io/compare/whisper-vs-parakeet Published: 2026-10-21 Updated: 2026-10-20 Comparison Two open-weight model families dominate local [speech recognition](https://metavert.io/speech-recognition) in 2026. [OpenAI](https://metavert.io/openai)'s **Whisper**, released in 2022 and last updated with large-v3 (2023) and large-v3-turbo (2024), is the default choice for multilingual transcription: one model, 99 languages, a permissive license and a huge ecosystem of ports such as whisper.cpp. [NVIDIA](https://metavert.io/nvidia)'s **Parakeet TDT 0.6B** is the newer challenger: a 600-million-parameter transducer model released as v2 (English, May 2025) and v3 (25 European languages, August 2025) that is markedly faster and, on English benchmarks, more accurate. The choice matters most when transcripts drive something downstream, such as captions, [transcript-based editing](https://metavert.io/transcript-based-editing) or [speaker diarization](https://metavert.io/speaker-diarization). Then word timing and behavior on silence matter as much as raw accuracy, and that's where the two architectures differ most. This comparison uses the Open ASR Leaderboard as of October 17, 2026, the model cards, published research on Whisper's hallucinations, and one dated, first-party local test. Both families are [open-weight models](https://metavert.io/open-weight-models) you can run on your own hardware. ## Feature Comparison | Dimension | Whisper | Parakeet TDT | | --- | --- | --- | | Maker and license | OpenAI; MIT for code and weights per the GitHub repo (Hugging Face lists large-v3 as Apache-2.0) | NVIDIA; CC-BY-4.0, so attribution is required | | Latest open weights | large-v3 (2023) and large-v3-turbo (2024) | TDT 0.6B v2 (May 2025) and v3 (August 2025) | | Parameters | 1.55B (large-v3); 809M (turbo) | 600M | | Languages | 99, plus translation into English (not in turbo) | v2: English; v3: 25 European languages with language detection | | Architecture | Encoder-decoder transformer over 30-second windows, decoding text token by token | FastConformer encoder with a Token-and-Duration Transducer (TDT) decoder | | Avg English WER, Open ASR Leaderboard (Oct 17, 2026) | 7.16% (large-v3); 7.64% (turbo) | 6.02% (v2); 6.18% (v3) | | Throughput (leaderboard RTFx) | 470 (large-v3); 794 (turbo) | 5,988 (v2); 6,132 (v3) | | Word timestamps | Not native; derived from cross-attention alignment (DTW) or external forced alignment such as WhisperX | Native word, segment and character timestamps | | Long audio per pass | 30-second windows; long files are chunked | Up to 24 minutes with full attention; v3 up to 3 hours with local attention | | Behavior on silence and music | Documented hallucinations over non-speech audio | NVIDIA claims resilience to music and silence; no independent study found | | Common runtimes | openai-whisper, whisper.cpp (1.9.5, October 2026), WhisperX | NVIDIA NeMo; sherpa-onnx (1.13.8) for CPU and edge | ## Detailed Analysis ### Two Architectures, Two Failure Modes Whisper is an encoder-decoder transformer. The encoder reads a 30-second window of audio and the decoder writes the transcript token by token, like a language model conditioned on sound. That design is why it handles so many languages and tasks (transcription, translation into English and language identification) in one model. It is also why it can produce fluent text that wasn't said: a decoder that is asked to write something will sometimes write the most likely phrase even when the audio is silence or music. Large-v3-turbo keeps the large-v3 encoder but cuts the decoder from 32 layers to 4, trading a small amount of quality for much faster decoding. OpenAI's README notes that turbo doesn't translate. Parakeet TDT is a transducer. A FastConformer encoder processes the audio, and the TDT decoder emits tokens together with their durations, aligned to the audio frames. Because each token is tied to a position in the audio, word timestamps come out of the model directly, and the decoder has less room to invent text where no speech exists. The trade-off is scope: v2 is English-only and v3 covers 25 European languages, against Whisper's 99. ### Accuracy and Speed on the Open ASR Leaderboard On the Hugging Face Open ASR Leaderboard's default English view (version dated October 17, 2026, averaging 11 datasets), Parakeet TDT 0.6B v2 scores 6.02% average word error rate and v3 6.18%, against 7.16% for Whisper large-v3 and 7.64% for turbo. The best proprietary systems lead the board, and NVIDIA's larger Canary-Qwen-2.5B does slightly better than Parakeet at 5.94%. The speed gap is wider than the accuracy gap: the leaderboard's inverse real-time factor (RTFx) is about 6,000 for Parakeet against 470 for large-v3 and 794 for turbo. (NVIDIA's model cards report about 3,300 at batch size 128, so compare RTFx figures only within one source.) Two caveats narrow the picture. On the leaderboard's long-form tab the models are nearly tied: Parakeet v3 13.37%, turbo 13.60%, v2 13.88% and large-v3 13.92%. And on multilingual audio Whisper large-v3 often wins. It's ahead of Parakeet v3 in German (4.00% vs 4.12%), Italian, Spanish (3.35% vs 3.71%), Portuguese and Dutch (4.42% vs 7.53%), while Parakeet v3 is ahead in French (5.38% vs 6.24%). Parakeet's English advantage doesn't automatically carry over to other languages. ### Word Timing For captions and editing, where each word must be placed on the timeline, timing is often more important than WER. Parakeet returns word, segment and character timestamps natively. Whisper doesn't predict timestamps per word. The reference implementation has derived them since March 2023 by aligning cross-attention weights with dynamic time warping, whisper.cpp offers an option it labels experimental, and WhisperX adds forced alignment with a separate wav2vec2 model and voice activity detection. In OpenAI's hosted API, word-level timestamps are supported only on whisper-1, which OpenAI has scheduled for shutdown on February 26, 2027; its recommended transcription model is now gpt-transcribe. ### Hallucinations and Silence Whisper's hallucinations are well documented. The "Careless Whisper" study (FAccT 2024) found that 1.4% of 13,140 audio segments produced hallucinated text, and that 38% of those hallucinations contained harmful content such as invented violence or false authority. Speakers with aphasia, whose speech has longer pauses, triggered them more often. A 2025 ICASSP paper showed that non-speech audio induces Whisper hallucinations and proposed filtering them out. NVIDIA says Parakeet is resilient to music and silence, but that is a vendor claim; no independent study of Parakeet hallucination was found. The practical mitigations for Whisper are voice activity detection before transcription and not conditioning each window on the previous text (see [AI hallucinations](https://metavert.io/ai-hallucinations)). ### What One Local Bake-Off Found MetaRush's local transcription bake-off (October 17, 2026, on an Apple M2 Max with 64 GB of RAM) compared whisper.cpp 1.9.5 running large-v3-turbo on Metal against Parakeet TDT 0.6B v2 and v3 (int8) running through sherpa-onnx 1.13.8 on the CPU, on synthetic fixtures and public-domain speech. These are single-machine measurements, not a benchmark, but they illustrate the trade-offs above: - **Accuracy:** Parakeet v2 had 0.8% WER on a 199-second public-domain speech recording and 0.6% and 0.9% on 7.9- and 31.5-minute mixes. Whisper turbo, with conditioning on previous text turned off, scored 2.4%, 1.1% and 1.3%. With default settings, Whisper fell into a repetition loop at the end of the 7.9-minute mix. - **Silence:** Whisper output "Thank you." over 8 seconds of digital silence and over a 30-second music bed, and only its built-in voice activity detection stopped that, at the cost of breaking word timestamps. Parakeet output nothing. - **Timing:** across an inserted 1.5-second pause, Parakeet v2 reported a 1.28-second gap; Whisper's default token times reported none and placed the next word about a second early, inside the pause. Turn-start error averaged 0.08 seconds for Parakeet v2 and 0.27 for Whisper. - **Parakeet v3 dropped sentences:** deterministically and depending on chunking, it lost 47 of 373 words in a single pass over the 199-second recording, and 63 of 965 words on the 7.9-minute mix in 30-second chunks (4 with 20-second chunks). v2 didn't. Neither the v3 model card nor the leaderboard shows this. - **Speed:** both were fast enough for local use, with real-time factors of about 0.02 for Whisper turbo on the GPU and 0.03 for Parakeet v2 on the CPU alone. In this ONNX build, Parakeet failed on a single 7.9-minute input, so audio had to be chunked. The bake-off chose Parakeet v2 as the default for English and kept Whisper as the multilingual fallback. That pairing is a reasonable starting point for anyone building captions or editing tools. ### Licensing and Ecosystem Both licenses permit commercial use. Whisper's MIT license (the GitHub repo covers code and weights; Hugging Face lists large-v3 as Apache-2.0, also permissive) has no attribution requirement beyond keeping the notice. Parakeet's CC-BY-4.0 requires attribution to NVIDIA. NVIDIA's newer streaming model, parakeet-unified-en-0.6b (April 2026), uses the NVIDIA Open Model License instead, so check each model card. Whisper's ecosystem is far larger, with ports for almost every platform. Parakeet's runs mainly through NVIDIA NeMo on GPUs and sherpa-onnx on CPUs and edge devices, which makes it practical for [edge AI](https://metavert.io/edge-ai) deployments too. ## Best For #### Languages Beyond Europe Whisper Whisper covers 99 languages; Parakeet v3 covers 25 European languages and v2 English only. #### English Captions and Transcript Editing Parakeet Native word timestamps, a lower English WER and clean behavior on silence make v2 the stronger choice for timed text. #### European Languages Tie Per-language results are mixed: Whisper large-v3 leads in German, Spanish and Dutch, Parakeet v3 in French. Test on your own audio. #### Translating Speech Into English Whisper Whisper large-v3 can translate into English directly (turbo can't); Parakeet only transcribes. #### High-Volume Batch Transcription Parakeet Roughly eight to thirteen times the leaderboard throughput of Whisper at a lower English error rate. #### Audio With Music and Long Pauses Parakeet Whisper's hallucinations over non-speech are well documented; if you use Whisper, gate it with voice activity detection. #### Running Locally on a Laptop Tie whisper.cpp on a GPU and Parakeet through sherpa-onnx on a CPU both run well faster than real time. ## The Bottom Line Choose **Parakeet TDT 0.6B v2** for English transcription where timing matters: captions, transcript-based editing, diarization and search. It's faster and more accurate on English benchmarks, returns native word timestamps and is less prone to inventing text over silence. Treat **v3** with care: its multilingual coverage is useful, but test it on your own long-form audio, because at least one local test found it dropping whole sentences. Choose **Whisper** when you need breadth: 99 languages, translation into English, the largest tooling ecosystem and a license without attribution requirements. Pair it with voice activity detection and an alignment step if you need reliable word timing, and plan around OpenAI's hosted whisper-1 shutdown in February 2027. For many products the right answer is both: Parakeet for English, Whisper as the multilingual fallback. For hosted alternatives, see [Deepgram vs AssemblyAI](https://metavert.io/compare/deepgram-vs-assemblyai). ## Related Topics - [Speech Recognition (ASR)](https://metavert.io/speech-recognition) - [Speaker Diarization](https://metavert.io/speaker-diarization) - [Auto-Captions](https://metavert.io/auto-captions) - [Transcript-Based Editing](https://metavert.io/transcript-based-editing) - [Speech and Voice AI](https://metavert.io/speech-voice-ai) - [Open-Weight Models](https://metavert.io/open-weight-models) - [NVIDIA](https://metavert.io/nvidia) - [OpenAI](https://metavert.io/openai) - [Deepgram vs AssemblyAI](https://metavert.io/compare/deepgram-vs-assemblyai) ## Further Reading - [OpenAI Whisper on GitHub](https://github.com/openai/whisper) — model sizes, languages, license - [NVIDIA Parakeet TDT 0.6B v2 model card](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2) — English; timestamps; CC-BY-4.0 - [NVIDIA Parakeet TDT 0.6B v3 model card](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3) — 25 European languages; long-form attention - [Hugging Face Open ASR Leaderboard](https://huggingface.co/spaces/hf-audio/open_asr_leaderboard) — WER and RTFx, English, long-form and multilingual views - [Careless Whisper: Speech-to-Text Hallucination Harms (FAccT 2024)](https://arxiv.org/html/2402.08021v2) — the hallucination study - [whisper.cpp releases](https://github.com/ggml-org/whisper.cpp/releases) — v1.9.5, October 2026 - [OpenAI API deprecations](https://developers.openai.com/api/docs/deprecations) — whisper-1 shutdown on February 26, 2027