Best local speech-to-text models for streaming

Live transcription is a latency problem, not just an accuracy problem. Native streaming, partial-result stability, backend requirements and language coverage matter more than a generic benchmark rank.

Native streaming versus chunked transcription

A model can appear to “stream” even if the underlying model repeatedly transcribes rolling windows. That can be useful, but it has different latency, compute and text-stability behavior from an architecture designed to cache state while new audio arrives. For a voice agent, dictation tool or live captioner, that distinction can matter more than a small difference in offline WER.

Nemotron 3.5: the clearest streaming-first candidate

NVIDIA's Nemotron 3.5 ASR Streaming 0.6B is explicitly a cache-aware streaming model. The current release also supports Transformers and exposes a Q8 GGUF in addition to NeMo assets. It is the first model to test when native streaming is non-negotiable and your locale is in a ready-to-use support tier.

The main caution is language support. NVIDIA publishes tiered locales rather than one flat list. An adaptation-ready locale can require fine-tuning and should not be treated as a transcription-ready deployment.

VibeVoice-ASR Streaming: native speaker-attributed live transcription

Microsoft released VibeVoice-ASR-Streaming in September 2026 as a distinct streaming model rather than a chunking wrapper around the long-form checkpoint. The current 1.5B-named model card reports 3B total parameters and lists ten languages: Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish.

Its differentiator is not merely low latency. It continuously emits who said what as audio arrives and accepts customized hotwords. For meetings, call transcription and multi-speaker live captions, that makes it a fundamentally different candidate from a plain transcript-only streaming model.

The published streaming card does not advertise full When timestamps in the same way the long-form VibeVoice-ASR card does, so do not assume the streaming output is a drop-in subtitle-timing pipeline. Test the exact output schema you need.

Qwen3-ASR: official streaming inference, with an important implementation caveat

Qwen3-ASR 0.6B and 1.7B support offline and streaming inference. The official project currently lists 30 languages plus 22 Chinese dialects. Streaming is available through the vLLM backend, but the current implementation accumulates audio from the beginning and re-feeds the audio seen so far for each update. It also does not return timestamps in streaming mode. If you need word-level timing, Qwen provides a separate 0.6B forced aligner for 11 languages.

This still makes Qwen3-ASR useful when you want one ASR family for batch and interactive transcription, but it should not receive the same native-streaming label as a cache-aware streaming encoder. Long-session latency and compute growth deserve explicit benchmarking.

Moonshine: optimized around live voice and edge hardware

Moonshine's current English streaming lineup spans Tiny Streaming at 34M parameters, Small Streaming at 123M and Medium Streaming at 245M. The framework caches work during speech and is designed for low response latency after a phrase ends. It also targets Python, Windows, macOS, Linux, iOS, Android and Raspberry Pi-class devices.

For English voice interfaces, that combination makes Moonshine a serious first benchmark. For other languages, check the exact model and license: Moonshine states that English-language models are MIT while other-language models use a non-commercial community license.

Where Whisper and Parakeet fit

Whisper has excellent runtime coverage and many applications stream it successfully with rolling windows, VAD and chunking. That is valuable engineering, but plain Whisper is not a native streaming architecture. Parakeet v3 is similarly better thought of as a high-throughput transcription model than a streaming-first model.

If your product is mostly recordings with an optional live preview, those families may still be the better overall choice. If every 100–300 milliseconds of response time matters, benchmark them against a native streaming family instead of relying on feature labels.

What to measure in a live test

• First partial latency: how quickly useful text appears after speech starts.
• Finalization latency: how long the final transcript takes after the speaker stops.
• Revision rate: how often partial text changes enough to feel distracting.
• Real-time compute load: whether the model can stay ahead of incoming audio on the actual device.
• Long-session stability: memory growth, VAD behavior and transcript drift over 30–60 minutes.
• Language and punctuation quality: especially when automatic language detection is enabled.

Bottom line

For local live transcription, start with the architecture instead of the leaderboard. Test Nemotron 3.5 for streaming-first multilingual work, Moonshine for English edge/live applications, and Qwen3-ASR when its official vLLM streaming workflow and language set fit your stack. Keep Whisper as the ecosystem baseline, and use Parakeet when the workload is more batch-heavy than interactive.

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.