2026 Local Speech Recognition Model Comparison: Qwen3-ASR, SenseVoice, Parakeet, Nemotron, Cohere, Whisper - Which to Choose?

If you just want the one-sentence conclusion: For Chinese or Chinese-English mixed speech, choose Qwen3-ASR 1.7B; for pure English offline high-accuracy transcription, prioritize Qwen3-ASR 1.7B, with Cohere Transcribe as a backup only for pure English batch processing without timestamps; for English real-time dictation, choose Parakeet Unified EN; for multilingual real-time subtitles or CPU deployment, choose Nemotron 3.5; for low-spec computers with pure Chinese, choose SenseVoice Small; for 99 languages and timestamps, consider Whisper large-v3-turbo.

This is not a cliché about "no best, only most suitable." Every recommendation below corresponds to specific public data, clear applicable conditions, and known limitations.

Quick Selection Table

Why This Comparison Differs from Common Leaderboards

Speech recognition model comparisons are prone to three major mistakes.

First, treating English leaderboards as universal capability benchmarks. Hugging Face's Open ASR Leaderboard primarily uses English short audio datasets. Being first on English mean doesn't mean first on Chinese, or first on Chinese-English mixed speech.

Second, directly comparing WER across different test sets. Chinese has no spaces for word segmentation, so we look at Character Error Rate (CER); English uses Word Error Rate (WER). AISHELL, WenetSpeech, FLEURS, and LibriSpeech differ in difficulty and accent distribution. Cross-dataset comparisons can only look at magnitude, not decimal points.

Third, treating vendor latency as your latency. "160 ms" or "100 ms" are usually results from optimal hardware, specific chunking, or even just time-to-first-token. Your end-to-end latency depends on quantization level, inference framework, CPU/GPU, and whether the model is already loaded.

Data in this article is current as of October 19, 2026, with sources in three tiers:

• Tier A: Official model cards, original papers, official release notes, Open ASR Leaderboard original records.
• Tier B: Third-party independent benchmarks with complete testing conditions.
• Tier C: Community measurements and media reports, used only to identify issues and support trends.

This article did not conduct new laboratory measurements. All numbers come from public sources, with verifiable links and a separate source ledger at the end. The benefit is you can verify them yourself; the drawback is testing conditions vary across sources, so this article doesn't publish a "total score ranking," only scenario-based judgments.

Six Models at a Glance

* SenseVoice's Q5 size comes from third-party conversion (FP16 conversion version ~0.45 GB); other Q5 sizes come from Voconly's own voconly-org GGUF conversion repository, not a neutral third-party measurement. File size doesn't equal peak memory; actual usage depends on inference framework and concurrency.

English Accuracy: How to Read the Open ASR Leaderboard

For English short audio, the Open ASR Leaderboard is currently the most comparable public benchmark. Below are numbers from official submissions or model card records:

This table must be read with dates. When Cohere launched on 2026-03-26 with 5.42, it topped the leaderboard. But by 2026-09-25, a third-party Open ASR snapshot showed Cohere had dropped to ~5.40, ~20th place (among 69 English models), while Qwen3-ASR-1.7B reached ~4.95, becoming the best open-source model in that snapshot. Hugging Face's results dataset for the second half of 2026 also showed Qwen3 (5.75) leading Cohere (6.26). These absolute values differ due to test sets and text normalization changes and can't be compared across snapshots, but the direction is consistent: "Cohere first in English" only held for the release day snapshot and shouldn't be used as a 2026-10 selection criterion.

This table also shows one thing: magnitude matters more than decimal points. Cohere, Qwen3, and Parakeet historical records all fall in the 5–6% range, Whisper turbo in 7–8%; snapshots can differ by 1–2 percentage points. For English dictation, you should look more at latency, language coverage, and hardware than obsessing over 0.1 percentage points in some snapshot.

But English mean also has blind spots. Third-party testing organization InferenceBench tested Cohere Transcribe: on LibriSpeech clean, WER was only 1.46%, but on AMI meeting audio, it reached 17.7%–29.2%. This shows "clean read audio" and "real meetings, accents, noise" can differ by orders of magnitude. Leaderboard scores are a starting point, not an endpoint.

Chinese Accuracy: No Unified Leaderboard, But Clear Conclusions

Chinese currently has no widely-cited unified leaderboard like Open ASR Leaderboard. Below are Chinese CER from various public sources. Different rows use different test sets and cannot be directly ranked; look at magnitude and trends.

There are three conclusions.

First, for Chinese, Qwen3-ASR 1.7B is the first choice. It falls in the 2.4–5.9% range across AISHELL-2, WenetSpeech, SpeechIO, and FLEURS-zh, the only one among the six with complete numbers across multiple Chinese benchmarks. Streaming mode FLEURS-zh is 2.84%, not much higher than offline 2.41%.

Second, SenseVoice Small remains the optimal choice for low-spec Chinese. Official only released results as charts, with wide variation in third-party conversion metrics (AISHELL-1 ~3.4–5.8%), but the trend is consistent: Chinese clearly stronger than Whisper, yet model size is only 234M. It's not suitable for long audio and real-time streaming, but excellent for short-sentence dictation on low-spec computers.

Third, Nemotron 3.5 is not a Chinese dictation model. Its Mandarin FLEURS CER is 19–21%, and official explicitly places Mandarin in "broad-coverage" rather than "transcription-ready" tier. Treating it as the first choice for Chinese real-time is a misreading. Its value lies in multilingual real-time and pure CPU deployment, not Chinese accuracy.

Cohere Transcribe's English mean is impressive, but public materials provide no Mandarin CER and explicitly state the model targets "single, preset language" with unstable performance on code-switched audio. Chinese users shouldn't choose it just because it ranked first on an English leaderboard.

Chinese-English Mixed and Multilingual: Don't Confuse "Number of Languages Supported" with "Code-Switch Capability"

"Supports 14 languages" and "can handle Chinese-English mixed speech" are two different things.

• Qwen3-ASR covers 30 languages and 22 Chinese dialects, making it the most suitable for Chinese-English mixed speech among these models. Its training objective and language coverage are more oriented toward multilingual scenarios, with Chinese dialects explicitly listed as a capability.
• SenseVoice Small supports Chinese, Cantonese, English, Japanese, and Korean. Chinese is strong, but English terms and long English passages are weaker than Qwen3; it's "five-language" rather than "code-switch specialized."
• Cohere Transcribe while supporting Chinese, requires specifying one language per request, with no automatic language identification; official admits code-switch performance is unstable. Should be directly excluded for Chinese-English mixed speech.
• Nemotron 3.5 has language ID condition and can automatically identify languages, but its design is "one audio, one main language," not seamless switching between Chinese and English within one sentence; Mandarin itself is only broad-coverage level.
• Whisper large-v3-turbo has the broadest language coverage, but no specific optimization for code-switch, and Chinese accuracy is clearly behind Qwen3.
• Parakeet Unified EN only supports English, directly inapplicable for code-switch scenarios.

If your work frequently involves sentences like "Can we move this deadline back a bit?" or "Need to push down the API latency a bit more," Qwen3-ASR is the more stable choice.

Latency: Four Different Metrics, Don't Mix Them

Before discussing "how fast," distinguish four metrics:

1. Time to First Token (TTFT): How long after you finish the first word before text appears.
2. time-to-final: How long for the final result of a sentence to stabilize.
3. End-to-end latency: Time from stopping speech to text being inserted at cursor.
4. RTF / RTFx: Ratio of processing time to audio duration, usually measured under high concurrency and batch processing.

Note that Qwen3's two numbers aren't from the same concurrency point: 92 ms is minimum TTFT at concurrency 1; RTF 0.064 is per-request RTF at concurrency 128, aggregate throughput ~2000× (processing ~2000 seconds of audio per second), while average TTFT at that concurrency rises to 3.21 s. Treating 0.064 as aggregate RTF underestimates throughput; treating 92 ms as high-concurrency latency underestimates wait time.

For speech input tools, end-to-end latency is the metric users can perceive. Vendor "160 ms" or "100 ms" are usually first-token or stabilized latency under optimal configuration; adding audio buffering, inference, post-processing, and cursor insertion makes actual experience higher. When evaluating or selecting, measure p50/p95 on your own machine rather than just looking at marketing pages.

Size, Quantization, and Hardware: Q5 File Size ≠ Memory Usage

Several important reminders:

• Q5 is usually Voconly's default download tier (check settings page for specifics), but public leaderboards mostly use FP16 or BF16. The two aren't directly equivalent.
• Size data is from product's own sources. The Q5 file sizes in the table above come from Voconly's voconly-org conversion repository, suitable for magnitude comparison, but not neutral third-party measurement; check repository file pages and settings page displays before downloading.
• Quantization loss varies by model. Third-party transcribe.cpp tests show Parakeet Unified EN's Q5_K_M scored 1.58% WER on LibriSpeech test-clean, basically matching NVIDIA official FP16's 1.63%; SenseVoice's Q5 port scored 3.18%. But this only applies to these two models on that test set and can't be extrapolated to Qwen3 or Cohere.
• File size isn't peak memory. Inference framework, context cache, batch size, and concurrency all significantly affect RAM/VRAM. Cohere 2B and Qwen3 1.7B actual usage is noticeably higher than 0.6B models.
• If your machine only has CPU, prioritize SenseVoice, Parakeet, Nemotron; Qwen3 1.7B and Cohere 2B are better suited for GPU or strong CPU environments.

Model-by-Model Review

Qwen3-ASR 1.7B: The Default Answer for Chinese and Chinese-English Mixed

Advantages: Most complete Chinese benchmarks, AISHELL-2 2.71, WenetSpeech 4.97/5.88, FLEURS-zh 2.41; streaming FLEURS-zh 2.84, minimal drop; high ranking on English leaderboards; 52 languages/dialects; Apache 2.0; timestamps can be supplemented with separate ForcedAligner.

Limitations: Q5 file ~1.41 GB, heavier than 0.6B models; streaming inference officially only supports vLLM, and streaming mode doesn't support batch processing and returning timestamps; public README doesn't list hotword/context biasing as basic functionality, see API or technical report for implementation.

Suitable for: Chinese dictation, Chinese-English mixed speech, multilingual office work, users needing high Chinese accuracy.

Not suitable for: Low-end CPU only, pursuing extreme lightweight, or must use non-vLLM backend for streaming.

SenseVoice Small: Optimal Solution for Low-Spec Chinese

Advantages: 234M parameters, third-party Q5 conversion ~164 MB; Chinese stronger than Whisper; in third-party tests RTF ~0.007, 10-second audio ~70 ms; supports Chinese, Cantonese, English, Japanese, Korean; FunASR pipeline can provide timestamps.

Limitations: Non-streaming, single session max 30 seconds; English and professional terms weaker than Chinese; official only publishes benchmarks as charts, third-party quantization/conversion metrics vary widely (AISHELL-1 ~3.4–5.8%); weights use FunASR Model License, not Apache/MIT.

Suitable for: Low-spec computers, pure Chinese short sentences, not requiring real-time performance, wanting smallest possible model.

Not suitable for: Long audio, real-time subtitles, Chinese-English mixed speech, teams needing unified commercial license review.

Cohere Transcribe: Top-Tier English Offline Accuracy, But Not an All-Rounder

Advantages: 2B parameters, at release Open ASR English mean 5.42 and topped leaderboard; 14 languages; Apache 2.0; long-audio chunking and batch processing design mature.

Limitations: Must manually specify language, no automatic language identification; doesn't support timestamps and speaker diarization; official admits code-switch is unstable; on real meeting audio, third-party tests show WER can reach 17.7%–29.2%; no official Chinese benchmark. Q5 file ~1.65 GB, largest among the six.

Suitable for: Clean pure English audio, offline batch transcription, users prioritizing English accuracy over real-time performance.

Not suitable for: Chinese, Chinese-English mixed speech, real-time subtitles, scenarios needing timestamps or speaker diarization.

Parakeet Unified EN: Default Answer for English Real-Time Dictation

Advantages: 0.6B, English-specific optimization; offline Open ASR mean 5.91; 160 ms buffered streaming; native punctuation and capitalization; Q5 ~0.52 GB; NVIDIA Open Model License.

Limitations: English only; at 160 ms streaming WER rises to 8.44, explicit accuracy-latency trade-off; not a multilingual model; timestamp capability undocumented.

Suitable for: English meetings, English dictation, English real-time subtitles, users needing acceptable accuracy at low latency.

Not suitable for: Chinese, multilingual, scenarios needing word-level timestamps.

Nemotron 3.5 ASR: Engineering Choice for Multilingual Real-Time/CPU, Not a Chinese Dictation Model

Advantages: 0.6B; 40 language regions (32 out-of-box usable: 19 transcription-ready + 13 broad-coverage; 8 require fine-tuning); 80/160/560/1120 ms adjustable chunking; cache-aware streaming, official claims time-to-final can be below 100 ms; CPU is officially supported backend; native punctuation and capitalization; strong concurrency.

Limitations: Mandarin FLEURS CER 19–21%, officially listed as broad-coverage; automatic language identification slightly worse than manual specification; not code-switch specialized; license is OpenMDW-1.1 (NVIDIA Open Model Data License), some third-party platforms mislabel as NVIDIA Open Model License, check official model card before commercial use.

On CPU boundaries: NVIDIA official documentation lists CPU as optional backend, LiveKit also explicitly states CUDA is not required, so "multilingual + pure CPU" is technically feasible. But pure CPU real-time performance depends on chunking and implementation: third-party ONNX tests show 80 ms chunking RTF ~0.65 (can't run real-time), 320 ms ~0.26, 1120 ms ~0.11; to use CPU for real-time subtitles, usually need larger chunking or stronger CPU.

Suitable for: Multilingual real-time subtitles, low-power devices, CPU deployment (recommend 560 ms+ chunking), scenarios not requiring high Chinese accuracy.

Not suitable for: Chinese dictation, Chinese meeting records, scenarios needing high-accuracy proper nouns.

Whisper large-v3-turbo: Baseline for Ecosystem and Language Coverage

Advantages: 809M; 99 languages; MIT license; native timestamps; most mature ecosystem, most tooling and tutorials; ~8× faster than large-v3.

Limitations: English Open ASR mean 7.83, lower than new-generation models; Chinese AISHELL-1 8.64, WenetSpeech net 13.51, clearly behind Qwen3; no native streaming; ~6 GB VRAM (third-party reference); translation capability is not its strength.

Suitable for: Needing coverage of many minor languages, needing timestamps, relying on mature tooling, not sensitive to highest accuracy.

Not suitable for: Chinese-primary, pursuing highest English accuracy, needing true streaming.

Scenario Decision Tree

Five Common Misconceptions

Misconception 1: Old leaderboard first place is still first. Any "ranked X" must come with date and specific snapshot. Cohere also hasn't published Mandarin CER, and official explicitly states code-switch is unstable.

Misconception 2: More parameters means more accurate. Whisper turbo has 809M, but English mean 7.83, lower than 0.6B Parakeet (5.91) and 1.7B Qwen3. Architecture, training data, and task matching matter more than parameter count.

Misconception 3: Vendor latency is your latency. 160 ms, 100 ms are usually optimal hardware, specific chunking, first-token latency; end-to-end experience is also affected by quantization, inference framework, and system load.

Misconception 4: Quantization always causes significant accuracy drop. Parakeet's Q5 on LibriSpeech test-clean basically matches FP16; but this is model and test set specific, can't be extrapolated. The right approach is to test on your own audio.

Misconception 5: More languages supported means stronger code-switch capability. Cohere supports 14 languages but doesn't support automatic language identification, code-switch is unstable; Nemotron supports 40 language regions but Mandarin is only broad-coverage level. Language count doesn't equal code-switch capability.

How to Choose in Voconly

Voconly models can be switched anytime in settings, default download tier is usually Q5 (check settings page for specifics). Practical initial configurations:

• Chinese/Chinese-English mixed: Qwen3-ASR 1.7B.
• Low-spec computer, pure Chinese: SenseVoice Small.
• English real-time: Parakeet Unified EN.
• Pure English offline high accuracy: Qwen3-ASR 1.7B; Cohere Transcribe only as backup for pure English batch processing without timestamps.
• Multilingual real-time/CPU: Nemotron 3.5 ASR.
• Minor languages, timestamps: Whisper large-v3-turbo.

The model is only the first layer. Actual experience also depends on microphone, noise environment, your accent, professional terminology, and text post-processing. Suggest spending a week comparing 2–3 models on the same set of real audio before fixing default.

Limitations and Verification Methods

Limitations:

• This article didn't conduct first-party laboratory measurements, all numbers from public sources.
• Different sources have different test sets, languages, quantization levels, and hardware, can't be treated as strictly controlled horizontal experiments.
• Open ASR Leaderboard submission times and dataset versions change; this article provides both historical records and newer snapshots, not treating any single snapshot as permanent ranking.
• Vendor latency data is usually optimal configuration, can't be directly equated to Voconly's end-to-end experience.
• Q5 file sizes come from Voconly's own GGUF conversion repository, not neutral third-party measurement.

You can verify this article in three steps:

1. Select a piece of your own real audio covering your main scenario (pure Chinese, pure English, or Chinese-English mixed), at least 30 sentences.
2. Run 2–3 models on the same audio, same quantization level, and same machine; use jiwer to calculate WER/CER, and record p50/p95 latency.
3. Compare results with public data in this article. If there's a big discrepancy, prioritize your own tests, because your accent, microphone, and scenario are the ultimate variables.

Conflict of Interest Statement

Voconly includes and distributes several models mentioned in this article, so potential conflict of interest exists. We try to use only official model cards, public leaderboards, and reproducible third-party tests, and write out each model's limitations; this article received no paid ranking from any model vendor. If you find data errors, please point them out, we'll update and record the date.

Changelog

• 2026-10-19: Initial release. Data current as of 2026-10-19; Open ASR data from official submission records of each model, Chinese data from official model cards and third-party public evaluations.
• 2026-10-19 (Revised): Revised English ranking conclusions based on 2026-09 latest snapshots; added Nemotron CPU applicability boundaries and language tiers, Qwen3 efficiency metrics, Cohere RTFx, Parakeet low-latency suggestions, and Q5 size source notes; corrected Nemotron license to OpenMDW-1.1.

FAQ

Which ASR model for Chinese voice input?

Prioritize Qwen3-ASR 1.7B. Its public CER on AISHELL-2, WenetSpeech, SpeechIO, FLEURS-zh all fall in 2.4–5.9% range. When machine performance is limited, choose SenseVoice Small.

Which for Chinese-English mixed speech?

Qwen3-ASR 1.7B. Cohere Transcribe requires manually specifying a single language and official admits code-switch is unstable; Nemotron 3.5 supports automatic language identification but Mandarin accuracy is only broad-coverage level.

Which for English real-time dictation?

Parakeet Unified EN. It provides 160 ms buffered streaming, offline Open ASR mean 5.91. If you also need multilingual or pure CPU, choose Nemotron 3.5, but know Chinese accuracy will drop significantly.

Is Nemotron 3.5's Chinese really that bad?

Its Mandarin FLEURS CER is 19–21%, officially listed as broad-coverage rather than directly transcription-ready. Not recommended for daily Chinese dictation; multilingual real-time subtitles and low-power CPU scenarios are its strengths.

Can Nemotron really run multilingual on pure CPU?

Yes. NVIDIA official documentation lists CPU as optional backend, LiveKit also explicitly states CUDA is not required. But pure CPU real-time performance depends on chunking and implementation: 80 ms chunking can hardly run real-time, 560 ms+ is more realistic; on low-power devices suggest testing p50/p95 with your own audio first.

Is Cohere still first in English? Why not recommended for Chinese?

No. Any ranking needs date and specific snapshot. Cohere is still suitable for pure English batch processing, but shouldn't be selected assuming "first." It also hasn't published Mandarin CER, and official explicitly states model targets single preset language, code-switch is unstable, no automatic language identification, no timestamps. For Chinese scenarios, choose Qwen3-ASR or SenseVoice.

How much accuracy does Q5 quantization lose?

Varies by model. In public third-party tests, Parakeet's Q5 on LibriSpeech test-clean basically matches FP16 (1.58 vs 1.63), SenseVoice Q5 on the same test set is 3.18; but this can't be extrapolated to Qwen3 or Cohere. The most reliable method is to test on your audio.

Which model is most resource-efficient?

SenseVoice Small is lightest (234M, Q5 ~164 MB); next are Parakeet and Nemotron (~0.5 GB). Qwen3 and Cohere's Q5 files are 1.4–1.7 GB, actual memory usage is higher.

Is Whisper still worth using?

Worth it, but depends on scenario. Its advantages are 99 languages, MIT license, native timestamps, and mature ecosystem; English and Chinese accuracy have fallen behind new-generation models. If you need minor language coverage or timestamps, Whisper is still practical.

Can these models be used commercially?

All six models offer commercially-usable license paths, but terms differ: Qwen3 and Cohere are Apache 2.0, Whisper is MIT, Parakeet is NVIDIA Open Model License, Nemotron is OpenMDW-1.1 (NVIDIA Open Model Data License), SenseVoice is FunASR Model Open Source License (commercial use must meet attribution etc. conditions). Nemotron's official model card explicitly states OpenMDW-1.1 applies; some third-party platforms mislabel as NVIDIA Open Model License, check official model card before commercial use.

Why no total score ranking?

Because English WER, Chinese CER, latency, language coverage, and licensing can't be weighted into a total score that holds for all users. Someone only doing Chinese dictation and someone needing 99-language subtitles have inherently different optimal choices.

Data Sources

• Qwen3-ASR GitHub (model variants, Chinese/English benchmarks, streaming and timestamp documentation)
• Qwen3-ASR-1.7B Model Card
• Qwen3-ASR Technical Report (arXiv:2601.21337)
• Cohere Transcribe Release Notes (English Open ASR 5.42 and 14 languages)
• Cohere Transcribe Documentation (no automatic language identification, no timestamps, single-language restriction)
• CohereLabs/cohere-transcribe-03-2026 Model Card
• NVIDIA Nemotron 3.5 ASR Streaming 0.6B Model Card
• NVIDIA Nemotron 3.5 ASR Transformers Documentation
• NVIDIA ACE Nemotron ASR Deployment Documentation (CPU/CUDA/Vulkan backends)
• LiveKit: Nemotron 3.5 ASR Multilingual Deployment (CUDA Not Required)
• transcribe.cpp: Nemotron 3.5 CPU/Q5 Tests
• NVIDIA Parakeet Unified EN 0.6B Model Card
• FunAudioLLM SenseVoice GitHub
• FunAudioLLM/SenseVoiceSmall Model Card
• OpenAI Whisper large-v3-turbo Model Card
• Hugging Face Open ASR Leaderboard
• Open ASR Leaderboard Results Dataset (including newer snapshots)
• Open ASR Leaderboard Paper (arXiv:2510.06961)
• Spokenly: 2026-09-25 Open ASR Snapshot Summary (third-party, Tier C)
• BELLE-2 Whisper Chinese Evaluation (AISHELL/WenetSpeech CER)
• InferenceBench: Cohere Transcribe Real Audio Evaluation
• transcribe.cpp: Parakeet Unified EN Q5 Evaluation
• transcribe.cpp: SenseVoice Small Q5 Evaluation
• voconly-org/Qwen3-ASR-1.7B-gguf
• voconly-org/cohere-transcribe-03-2026-gguf
• voconly-org/parakeet-unified-en-0.6b-gguf
• voconly-org/nemotron-3.5-asr-streaming-0.6b-gguf
• voconly-org/whisper-large-v3-turbo-gguf

---

Download Voconly from GitHub →

Your voice. Your device. Your choice.

Last updated: October 19, 2026. Model performance and benchmark data change quickly; please refer to official model cards and leaderboards.