Best local speech-to-text models in 2026
There is no single best local ASR model. The useful answer depends on your language, hardware, whether you need true streaming, how important timestamps are, and what license obligations you can accept.
Quick comparison
| Family | Best reason to test it | Streaming | Language position | License note |
|---|---|---|---|---|
| Whisper | Mature ecosystem, many model sizes, broad runtime support | Usually wrapper/chunk based | Broad multilingual support | MIT |
| Parakeet TDT 0.6B v3 | Modern high-throughput transcription and timestamp support | Not a streaming-first architecture | 25 listed languages | CC BY 4.0 |
| Nemotron 3.5 ASR Streaming | Native low-latency streaming | Native | Tiered locale support | OpenMDW-1.1 |
| Qwen3-ASR | Official offline/streaming workflow and strong multilingual scope | Streaming via vLLM; current implementation re-feeds accumulated audio | 30 languages + 22 Chinese dialects | Apache-2.0 |
| Moonshine | Live voice and edge-device efficiency | Native streaming models | Language-specific models | English models MIT; other-language terms differ |
| Omnilingual ASR | Maximum language breadth | Not the main reason to choose it | 1,600+ language codes | Apache-2.0 |
1. Start with the constraint that can disqualify a model
Language support is the first hard gate. A model can be fast and accurate on a benchmark yet still be the wrong choice if the vendor does not list your target language or if the language is only adaptation-ready. After language, check deployment hardware, then latency requirements, timestamps, licensing and runtime maturity.
Do not rank models from a single public WER table and stop there. Error rate changes with language, accent, microphones, background noise, domain vocabulary, chunking strategy and decoding settings. A small benchmark on your own audio is more useful than a generic winner.
2. Whisper: the baseline that is hard to avoid
Whisper remains the most practical control model because it has multiple sizes from tiny through turbo and a very large local ecosystem. That makes it useful on CPU-only machines, Apple Silicon and NVIDIA GPUs. OpenAI describes turbo as an optimized large-v3 derivative that is much faster with minimal accuracy degradation, while the non-turbo multilingual models also retain Whisper's speech-translation workflow.
Whisper's weakness for this decision is not that it is obsolete; it is that many live implementations are wrappers around a model that was not designed as a native streaming architecture. If first-partial latency is the product requirement, compare it with a streaming-first family instead of assuming chunked Whisper is equivalent.
3. Parakeet: strong when batch throughput and timestamps matter
NVIDIA's Parakeet TDT 0.6B v3 is a 600M model with an explicit 25-language set. Its official repository now exposes Transformers, NeMo and a Q8 GGUF for local inference. It is a natural candidate for recordings, subtitles and pipelines where throughput matters more than native streaming behavior.
The trade-off is language coverage and licensing. If your language is outside the listed set, do not infer support from the word “multilingual.” CC BY 4.0 also carries attribution obligations that are different from MIT or Apache-2.0.
4. Qwen3-ASR and Nemotron: two different streaming approaches
Qwen3-ASR ships 0.6B and 1.7B models. The official project currently documents 30 languages plus 22 Chinese dialects, offline and streaming inference, long-audio transcription and a separate forced aligner for timestamps. Its vLLM streaming path currently re-feeds the accumulated audio seen so far and does not return timestamps, so we classify it as streaming via runtime rather than a cache-based native streaming architecture.
Nemotron 3.5 is more explicitly built around cache-aware low-latency streaming. Its locale support is tiered, which matters: a transcription-ready locale and an adaptation-ready locale are not interchangeable. If you need a model that works immediately without fine-tuning, verify the exact locale tier before you build around it.
5. Moonshine: a different optimization target
Moonshine is aimed at live voice applications and constrained hardware. Its current English streaming family includes Tiny Streaming, Small Streaming and Medium Streaming, while the project also publishes language-specific non-English models. The framework runs across desktop, mobile and edge platforms and includes higher-level voice features around transcription.
License terms differ by model language: the project states that English models are MIT, while other-language models use the Moonshine Community License and are non-commercial. That is exactly why a “family-level commercial” label is dangerous.
6. Omnilingual ASR: choose it when breadth is the point
Meta's Omnilingual ASR covers more than 1,600 language codes and offers multiple CTC and LLM-ASR checkpoints. The 300M-class CTC model is the sensible lightweight starting point in this matcher. Meta also publishes unlimited-audio variants for part of the family.
Extreme coverage is valuable, especially for under-served languages, but it is not a promise of uniform accuracy. Treat Omnilingual ASR as a candidate generator: if it covers a language that narrower families do not, benchmark the exact checkpoint on real data.
7. VibeVoice and Fun-ASR: two 2026 additions worth benchmarking
Microsoft VibeVoice-ASR changes the meeting-transcription decision because the base model can produce speaker-attributed, timestamped Who / When / What output for recordings up to 60 minutes in a single pass. That can remove an entire diarization/alignment stage when the workflow fits, but the full model is much heavier than Whisper, Parakeet or the 600M-class streaming models.
The September 2026 VibeVoice-ASR-Streaming release is a separate streaming family. Its 1.5B-named checkpoint reports 3B total parameters on the current model card, supports 10 languages and continuously emits speaker-attributed transcription. It is a more direct benchmark for live meeting transcription than forcing a batch model into small chunks.
Fun-ASR-MLT-Nano-2512 is a different proposition: 800M parameters, 31 explicitly listed languages and Apache-2.0. It is attractive as a compact multilingual challenger, but the current model card still lists timestamp return and speaker diarization as TODO items, so subtitle and meeting workflows may need extra components.
How to benchmark the shortlist
- Collect at least 20–30 minutes of representative audio, not a clean demo clip.
- Include accents, noise, overlapping speech, names and domain vocabulary you actually care about.
- Measure word or character error rate where appropriate, but also measure real-time factor, first-partial latency and peak memory.
- Run long files, not just short utterances, if meetings or media are your target workload.
- Verify timestamp behavior and speaker separation separately; they may come from a different component than the ASR model.
- Re-check the exact model license before redistribution or commercial deployment.
Bottom line
If you want one safe starting experiment, benchmark Whisper plus one model optimized for your dominant constraint. Pair Whisper with Parakeet for batch transcription, Nemotron for native multilingual streaming or Qwen3-ASR when its vLLM streaming workflow fits, Moonshine for English edge/live speech, or Omnilingual ASR for maximum language coverage. The best local STT model is the one that survives your language, hardware and workload tests.
Primary sources
Reviewed against primary sources on September 13, 2026. Model behavior, runtime support, language coverage and licenses can change; re-check the linked primary source before production use.