SMSpeech Model Match
Home / Models
Reviewed directory

Local speech-to-text model directory

Model-level facts and license notes used by the matcher. This is not a universal benchmark ranking.

Hardware fit in the interactive matcher is deliberately qualitative. Exact speed and memory use depend on runtime, quantization, audio, drivers and hardware.
Turn the directory into a decision

Do more than scan 20 rows.

Search the reviewed facts below, compare a shortlist side by side, or jump back to the matcher if you want the site to rank candidates for your exact constraints.

Get an automatic shortlistBest when you know your language, hardware and main workload but not the model family. Compare 2–3 modelsBest when you already have names and want documented capabilities side by side. Benchmark your finalistsBest when you have actual model outputs and want WER, speed and memory in one workspace.

Which of these do people download most? See the live popularity tracker.

Find a model in the directory

Search across model names, vendors, language coverage, streaming notes and licenses without scanning the full table.

26 models shown

Combine these filters with the search box—for example, native streaming + permissive license + multilingual.

Comparison workbench

Compare models side by side

Choose two or three reviewed models to compare their documented language coverage, streaming path, timing support and license. The comparison runs in your browser and uses the same source-reviewed data as the matcher.

Choose two or three models to compare.

Documented differencesCompare the trade-offs, then benchmark the finalists.
Local speech-to-text model comparison directory
ModelLanguage coverageStreamingLicenseWhat to knowPrimary source
Whisper tiny
OpenAI · 39M
99+ listed languagesWrapper / chunkedMITVery small multilingual baseline. Good when memory and CPU load matter more than maximum accuracy.Official source ↗
Whisper base
OpenAI · 74M
99+ listed languagesWrapper / chunkedMITLightweight multilingual option with a useful accuracy/speed step up from tiny.Official source ↗
Whisper small
OpenAI · 244M
99+ listed languagesWrapper / chunkedMITStrong general-purpose local choice when you want multilingual support without the footprint of larger Whisper models.Official source ↗
Whisper medium
OpenAI · 769M
99+ listed languagesWrapper / chunkedMITAccuracy-oriented multilingual Whisper choice. Better suited to capable local hardware than CPU-only systems.Official source ↗
Whisper turbo
OpenAI · 809M
99+ listed languagesWrapper / chunkedMITOptimized large-v3 derivative with much faster transcription and minimal accuracy degradation; not the choice for speech translation.Official source ↗
Parakeet TDT 0.6B v3
NVIDIA · 600M
25 European languagesWrapper / chunkedCC BY 4.0High-throughput multilingual ASR with 25 explicitly listed languages, word/segment timestamps and chunked streaming options. CC BY 4.0 attribution applies.Official source ↗
Orukeet
Oruk AI · 627M
25 European languagesNoCC BY-SA 4.0September 2026 Parakeet-derived 627M recognizer covering the same 25 listed languages. Official releases include NeMo, ONNX INT8 and native Q8/F16 paths with segment timing. The open local model is batch-first rather than native incremental streaming. Weights are CC BY-SA 4.0: attribution and applicable ShareAlike obligations apply.Official source ↗
Nemotron 3.5 ASR Streaming 0.6B
NVIDIA · 600M
40 language-locales in tiersNativeOpenMDW-1.1Native low-latency streaming with tiered locale support; some locales are adaptation-ready rather than out-of-the-box transcription-ready.Official source ↗
Qwen3-ASR 0.6B
Qwen · 0.6B
30 languages + 22 Chinese dialectsWrapper / chunkedApache-2.0Modern multilingual ASR with official vLLM streaming inference. The current streaming implementation re-feeds accumulated audio rather than using a cache-based streaming encoder, and it does not return timestamps. Detailed timestamps use the separate Qwen3 Forced Aligner.Official source ↗
Qwen3-ASR 1.7B
Qwen · 1.7B
30 languages + 22 Chinese dialectsWrapper / chunkedApache-2.0Larger Qwen3-ASR option for accuracy-oriented multilingual work. Official vLLM streaming inference re-feeds accumulated audio rather than using a cache-based streaming encoder, and detailed timestamps use the separate Qwen3 Forced Aligner.Official source ↗
Distil-Whisper large-v3
Hugging Face / Distil-Whisper · large-v3 distilled
EnglishWrapper / chunkedMITEnglish-only distilled Whisper model aimed at faster inference while retaining strong long-form transcription quality.Official source ↗
Moonshine Tiny Streaming (English)
Moonshine AI · 34M
EnglishNativeMITCompact English streaming ASR for low-latency and constrained-device use. Current Moonshine documentation lists Tiny Streaming at 34M parameters; the English model is MIT-licensed.Official source ↗
Moonshine Medium Streaming (English)
Moonshine AI · 245M
EnglishNativeMITEnglish streaming ASR for latency-sensitive local applications. Current Moonshine documentation lists Medium Streaming at 245M parameters.Official source ↗
Omnilingual ASR CTC 300M
Meta AI · 325M
1,600+ language codesNoApache-2.0Broad-coverage CTC ASR from Meta with a published 1,600+ language-code list. Coverage does not guarantee equal accuracy across languages, so benchmark representative audio before production.Official source ↗
Cohere Transcribe 03-2026
Cohere Labs · 2B
14 listed languagesNoApache-2.02B open-weights ASR model with strong published accuracy across 14 listed languages. The Hugging Face repository is gated even though the model is Apache-2.0.Official source ↗
VibeVoice-ASR 7B
Microsoft · 9B total
51 listed languagesNoMITLong-form model marketed as VibeVoice-ASR-7B; the current Hugging Face artifact reports 9B total parameters. Accepts up to 60 minutes in one pass and emits speaker-attributed, timestamped Who/When/What transcripts across 51 listed languages. MIT licensed; benchmark on your own meetings before production.Official source ↗
VibeVoice-ASR Streaming 1.5B
Microsoft · 3B total
10 listed languagesNativeMITNative streaming speaker-attributed ASR released in September 2026. The 1.5B-named variant reports 3B total parameters on Hugging Face and supports 10 listed languages plus customized hotwords. The model card advertises Who/What streaming output, not full timestamped When output.Official source ↗
VibeVoice-ASR-BitNet
Microsoft · 0.3B quantized
7 listed languagesWrapper / chunkedMITCPU-focused compressed VibeVoice-ASR variant: 1.58 GB quantized runtime, 7 listed languages and vendor-reported real-time throughput on 3+ CPU threads. Real-time throughput is not the same as a native streaming architecture, so it is classified as runtime/wrapper streaming here.Official source ↗
Fun-ASR-MLT-Nano-2512
FunAudioLLM · 800M
31 listed languagesWrapper / chunkedApache-2.0Compact 800M multilingual Fun-ASR checkpoint covering 31 listed languages with hotword and ITN controls. The current model card still lists timestamp return and speaker diarization as TODO items; streaming is treated as a toolkit/runtime workflow rather than a proven native streaming architecture.Official source ↗
Voxtral Mini 4B Realtime 2602
Mistral AI · 4B
13 listed languagesNativeApache-2.0Purpose-built 4B native-streaming ASR with configurable latency down to sub-200ms and 13 explicitly listed languages. Mistral releases the weights under Apache-2.0 and says the realtime path is not compatible with diarization. The official repository contains individual 8.86 GB weight files, so plan for capable local hardware and benchmark the exact latency setting you intend to use.Official source ↗
Granite Speech 5.0 2B
IBM · 2B
6 listed languagesNoApache-2.0IBM's October 2026 speech model with Apache-2.0 weights, six listed languages and speech translation to and from English. It transcribes segments of up to about two minutes, so split longer audio first. Documented runtimes include Transformers, vLLM, llama.cpp and mlx-audio. Ratings are provisional until independent benchmarks are available.Official source ↗
Granite Speech 5.0 470M TurboCTC
IBM · 470M
7 listed languagesNoApache-2.0Compact non-autoregressive CTC model that IBM positions for laptops, smartphones and other edge devices. Seven listed languages, Apache-2.0, native in Transformers 5.16 and available through mlx-audio on Apple Silicon. Ratings are provisional until independent benchmarks are available.Official source ↗
VibeVoice-ASR-Streaming-7B
Microsoft · 9B total
10 listed languagesNativeMITLarger sibling of VibeVoice-ASR Streaming 1.5B: native streaming who-said-what transcription in the same ten languages, with customized hotwords. The model card lists 9B parameters and advertises Who/What output rather than timestamped When output. Ratings are provisional until independent benchmarks are available.Official source ↗
Parakeet Redux
Moondream · 1.58-bit Parakeet 0.6B (178 MB)
25 European languagesNoCC BY 4.0A 1.58-bit (ternary) version of Parakeet TDT 0.6B v3 that fits in 178 MB and keeps its 25 languages and CC BY 4.0 license. Moondream reports 113× real time on eight x86 CPU cores with its Photon runtime; streaming and timestamps are not stated on the card. Ratings are provisional until independent benchmarks are available.Official source ↗
Whistle
Cactus Compute · 16.9 MB file (parameters not stated)
7 listed languagesNativeApache-2.0A single 16.9 MB streaming speech-to-text model for phones, wearables, robots, smart-home devices and microcontrollers. Seven listed languages, word-level timestamps, CPU-only with no GPU or dependencies, Apache-2.0. Ratings are provisional until independent benchmarks are available.Official source ↗
Jais-ASR
IFM · 5.7B total
Arabic (MSA + 19 dialects), plus English, French, Spanish, ItalianNoApache-2.0Arabic specialist covering Modern Standard Arabic and dialects from 19 Arab countries, with code-switching into English, French, Spanish and Italian. Apache-2.0; plan on roughly 12 GB of GPU memory in BF16, with FP8 and 4-bit AWQ builds also published. Ratings are provisional until independent benchmarks are available.Official source ↗
How to read the directory

Model facts are not a leaderboard

The directory separates documented capabilities from our qualitative fit scoring. Use it to build a shortlist, then benchmark the finalists on your own audio.

Language coverage

“Multilingual” is too vague to be useful. We store explicit language lists where vendors publish them and distinguish ready-to-use support from adaptation-only coverage.

Streaming labels

Native means the family exposes a streaming-first path. Wrapper means live behavior is typically created by chunking or runtime logic around the model. Those can both work, but they are not the same architecture.

License notes

MIT, Apache-2.0, CC BY 4.0 and custom model licenses carry different obligations. The directory surfaces the reviewed license but does not replace the current license text.

What to benchmark

Five checks before production

  1. Your exact language and accent: published coverage does not guarantee equal accuracy across languages.
  2. Long-form stability: run real meetings, interviews or media instead of only short clean clips.
  3. Latency and throughput: live apps care about first-partial/finalization delay; batch systems care about total processing speed.
  4. Timing and speaker separation: timestamps and diarization may come from separate components rather than the core ASR model.
  5. Runtime memory: checkpoint size is not peak VRAM or RAM. Measure the implementation you will actually ship.

Need a structured shortlist? Start with the 2026 local STT guide or use the interactive matcher.

Model profiles

Open a full profile

Each profile lists the reviewed languages, license, streaming and diarization support, hardware fit ratings and matcher ranks. Looking for a language instead? Use the language support matrix.