| Model | Type | Params / Size | Languages (CN/JP) | Window size | Speed | Response rate / latency | Accuracy | Key limitations | License | Free run path (no card) | Stack fit (A: Py·Flutter / B: Java·RN / Web) |
|---|---|---|---|---|---|---|---|---|---|---|---|
VibeVoice-ASR ★
frontierlong-form |
ASR (long-form, structured) | ~1.5B class 7.5 Hz tokenizer |
50+ langs CN JP |
Up to 60 min single-pass 🏆 no chunking needed |
Efficient (low frame-rate) | Batch long-form (not sub-second live) | High; outputs Who+When+What (speaker + timestamps) | New/early tooling; GPU for speed; long-form batch focus | Open (MIT) | Self-host; HF Transformers (since Mar 2026); HF Spaces | 🐍 A: native · ☕ B: via API/Py-sidecar · 🌐 Web: via backend |
Qwen3-ASR ★
frontierCN dialectsstreaming |
ASR + ForcedAligner | 1.7B & 0.6B 1.7B = Qwen3-1.7B + 300M AuT enc |
52 langs + 22 Chinese dialects CN★ JP |
Dynamic 1–8 s window streaming + offline long |
🚀 0.6B ≈ 2000 s audio / 1 s (concurrency 128) | TTFT ≈ 92 ms (excellent live) | High multilingual; LID + timestamps | Best speed needs GPU; newer ecosystem | Open source (Apache-2.0) | Self-host; HF; ModelScope | 🐍 A: native · ☕ B: API/sidecar · 📱 0.6B edge-capable · 🌐 ONNX possible |
Voxtral (Mini / Small / Realtime) ★
frontiermost accurate open |
ASR + translate + QA + summarize | Mini ~3B (Ministral 3B) Small 24B (Mistral Small 3.1) |
Many; Transcribe V2 = 13 langs w/ diarization CN JP (less CN-tuned) |
Long-form context | Fast; Mini = edge-friendly | Realtime: sub-200 ms live; word timestamps | 🏆 Voxtral Small = most accurate open-weights (AA-WER 2.9%) | Small (24B) heavy; CN/JP not its focus | Open weights (Apache-2.0) | Self-host; HF; voxtral.c (pure C, no Python) |
🐍 A: native · ☕ B: C/JNI or API · 📱 Mini/C edge · 🌐 via backend |
Kimi-Audio-7B-Instruct
frontierCN strong |
Audio foundation (ASR+AQA+TTS+dialogue) | 7B Qwen2.5-7B + Whisper enc |
Multilingual; CN★ JP | Standard (chunked) | Moderate (7B → GPU) | Conversational; not sub-second | Leading open-source ASR benchmarks (13M hrs trained) | Heavy (7B GPU); broad model, not edge | Open weights + training code | Self-host; HF | 🐍 A: native · ☕ B: API/sidecar · 📱 no (too heavy) |
GLM-4-Voice-9B
frontierCN+EN |
End-to-end speech LLM (voice chat) — ASR-capable | 9B 12.5 Hz tokenizer (from Whisper) |
CN★ + EN only | Conversational | Moderate (9B GPU); streaming decoder | Low-latency dialogue (starts at ~10 tokens) | Good (chatbot-oriented, not pure transcription) | Speech-dialogue focus, not best for plain transcripts; CN+EN only | Code Apache-2.0; weights = GLM Model License (free) | Self-host; HF | 🐍 A: native · ☕ B: API · 📱 no |
Qwen3-Omni-30B-A3B
frontieromni-modal |
Omni LLM (text+audio+image+video, real-time speech) | 30B MoE (~3B active) | Multilingual; CN★ JP | Long multimodal context | Real-time speech generation | Real-time capable (high-end GPU) | Very high (frontier omni) | Overkill for pure STT; needs strong GPU/VRAM | Open source (Apache-2.0) | Self-host; HF | 🐍 A: native · ☕ B: API · 📱 no |
Whisper (tiny→large-v3 / turbo) ★
99+ langsfree on Groq/HF |
ASR + translate | 39M / 74M / 244M / 769M / 1.55B; turbo 809M | 99+; CN JP (very good) | 30 s sliding (chunk longer) | tiny ≈10× large; turbo ≈8× large-v3 | Batch (live via WhisperX/streaming wrappers) | High (large-v3 best multilingual) | 30 s window → chunking; hallucinates on silence (use VAD) | Open (MIT) | Self-host (faster-whisper/whisper.cpp); Groq free; HF |
🐍 A: native · ☕ B: Groq API / whisper.cpp-JNI · 📱 ggml on-device · 🌐 transformers.js/WASM |
NVIDIA Canary-Qwen 2.5B
top EN accuracy |
ASR (LLM decoder) | 2.5B | English (Canary-1B: EN/DE/ES/FR + translate) | ~40 s | Slower (LLM decoder) | Batch | 🏆 ~5.63% WER (near top Open ASR Leaderboard) | English-centric; heavy; NeMo toolkit | Open (NVIDIA OpenModel / CC-BY) | Self-host (NeMo); HF | 🐍 A: native · ☕ B: API/sidecar · 📱 no |
NVIDIA Parakeet TDT 0.6B v2
fastest |
ASR (streaming-capable) | 0.6B / 1.1B | English (some multiling variants) | Streaming-capable | 🚀 RTFx > 2000 (≈6.5× Canary) | Real-time | High (EN); 10th on Open ASR Leaderboard | Mostly English; needs NeMo/GPU | Open (CC-BY) | Self-host (NeMo); HF | 🐍 A: native · ☕ B: sidecar · 📱 limited |
IBM Granite Speech 4.1 2B / 3.3 8B
SOTA accuracy |
ASR | 2B / 8B | English-focused (+ some) | Standard | Slower (8B) | Batch | 🏆 5.33% WER (4.1 2B = best open on Open ASR Leaderboard) | English-centric; heavy | Open (Apache-2.0) | Self-host; HF | 🐍 A: native · ☕ B: sidecar · 📱 no |
Meta MMS
1000+ langs |
ASR (massively multilingual) | ~1B | 1,000+ languages; CN JP | Standard | Medium | Batch | Good (esp. rare/low-resource langs) | Setup complexity; per-lang adapters | Open (CC-BY-NC for some / mixed) | Self-host; HF | 🐍 A: native · ☕ B: sidecar · 📱 no |
Meta SeamlessM4T v2
ASR+translate |
ASR + speech↔text translation | 2.3B | ~100; CN JP | Standard | Medium | Batch | Good; strong for translation | Large; GPU recommended; license restrictions | Open (CC-BY-NC — non-commercial) | Self-host; HF | 🐍 A: native · ☕ B: sidecar · 📱 no |
Wav2Vec 2.0 / HuBERT
fine-tunable |
ASR (self-supervised) | 95M – 1B | EN + any (fine-tune per lang) | Standard | Fast | Batch / streaming variants | Good when fine-tuned | No punctuation/casing by default; needs fine-tuning | Open (MIT / Apache) | Self-host; HF | 🐍 A: native · ☕ B: sidecar · 🌐 transformers.js |
Vosk ★
on-device |
ASR (streaming, offline) | ~50 MB small → 1.8 GB big | 20+; EN, CN, RU, ES, JP… | Streaming (zero-latency) | Real-time on CPU/phone | Live partials, fully offline | Small = medium; big = good | Per-language model; small models less accurate | Open (Apache-2.0) | Self-host / fully on-device (no server) | 🐍 A: native · ☕ B: native Java API · 📱 Android+iOS SDK · 🌐 WASM |
Moonshine (v2, 2026) ★
edge real-time |
ASR (streaming, edge) | Tiny / Base (small) | EN, ES, Mandarin CN, JP, KO, VI, UK, AR | Streaming | 🚀 Beats Whisper tiny/base; 5× faster on edge | < 200 ms latency | High for its size | Fewer languages than Whisper | Open (MIT) | Self-host / on-device (runs on MCU→phone→server) | 🐍 A: native · ☕ B: C/bindings · 📱 RN+Flutter on-device · 🌐 WASM |
Silero STT
lightweight |
ASR (streaming) | Small | EN, DE, ES, others | Streaming | Fast on CPU | Live | Medium | Fewer languages/updates | Open (CC-BY-NC / mixed) | Self-host / on-device | 🐍 A: native · 📱 mobile · 🌐 ONNX |
SenseVoice (Small) ★
CN+JP fast |
ASR + emotion + audio-event (non-autoregressive) | Small | CN★ Cantonese EN JP KO (50+) | Utterance-level (≤~30 s best) | 🚀 ~15× faster than Whisper-large | Low latency (batch + real-time) | Excellent CN, strong JP | Best on short clips; via FunASR | Open source | Self-host; HF; ModelScope | 🐍 A: native · ☕ B: sidecar · 🌐 ONNX |
Paraformer
Mandarin |
ASR (non-autoregressive) | Medium | CN★ Mandarin (EN variants) | Utterance / streaming | 🚀 Fast (parallel decode) | Streaming + batch | 🏆 Excellent Mandarin (60k hrs) | Mandarin-first; via FunASR toolkit | Open source | Self-host; ModelScope; HF | 🐍 A: native · ☕ B: sidecar · 🌐 ONNX |
FunASR (toolkit)
pipeline |
Framework: ASR + VAD + punctuation + diarization | — | CN / EN (hosts SenseVoice, Paraformer…) | Streaming + batch | Fast | Streaming | Excellent (CN) | It's a toolkit — pick a model inside | Open source | Self-host | 🐍 A: native · ☕ B: sidecar |
Kotoba-Whisper (v2 / bilingual) ★
JP best |
ASR (distilled Whisper large-v3) | distilled (~small) | JP★ + EN | 30 s | ~6× faster than large-v3 | Batch | 🏆 Excellent Japanese | JP/EN focus only | Open (MIT-style) | Self-host; HF | 🐍 A: native · ☕ B: sidecar · 📱 ggml possible |
ReazonSpeech
JP |
ASR (JP; JP→EN translate) | Medium | JP★ | Standard | Fast | Batch | Excellent Japanese | Japanese-centric | Open source | Self-host; HF | 🐍 A: native · ☕ B: sidecar |
Groq — Whisper large-v3 / turbo ★
no card to start |
Hosted ASR API (runs OSS Whisper) | n/a (you call API) | 99+; CN JP | File-based | 🚀 Very fast (LPU) | Free: ~2000 req/day, 7200 audio-sec/hr | High (Whisper large-v3) | Batch only; ~25 MB file cap; per-day quota | Model = MIT; service free tier | ✔ Free API, no credit card to start | 🐍 A: HTTP · ☕ B: HTTP (ideal — no AI libs) · 🌐 via backend |
Hugging Face Inference (Serverless)
no card |
Hosted endpoints for OSS models (Whisper, etc.) | model-dependent | model-dependent | model-dependent | Cold starts; variable | Small free monthly credits / rate-limited | model-dependent | Tight free limits; cold-start latency | varies by model | ✔ Free tier, no card to start | 🐍 A: HTTP · ☕ B: HTTP · 🌐 via backend |
Web Speech API
zero setup |
Browser-native ASR (not OSS, but free) | n/a | many (Chrome→Google) | Streaming | Real-time | Live partials, no key | Medium-good (varies) | Chrome sends audio to Google; inconsistent across browsers; not for private audio | Browser API (not open-source) | ✔ Free, no key, no card | 🌐 Web only (great Phase-0 demo); 📱 limited |
Every model here runs natively in Python (Hugging Face transformers or the model's own repo). This is the smoothest path for self-hosting. Flutter calls your Python API over HTTPS.
Java has weak native AI libs. Three integration paths: (1) call a free hosted API like Groq over HTTP (easiest), (2) run a tiny Python ASR microservice beside Spring Boot, or (3) use native bindings — Vosk has a Java API; whisper.cpp/voxtral.c via JNI.
Web never runs heavy models directly — it records audio and calls the backend. Exceptions that run in-browser: Web Speech API, and small models via transformers.js/ONNX/WASM (Whisper-tiny, Moonshine, Vosk-WASM).
For true offline mobile, use Vosk (Android+iOS SDK), Moonshine (RN/Flutter), or whisper.cpp ggml. Larger models (7B+) cannot run on phones — call the backend instead.
Groq Whisper large-v3 — fast, generous free tier, 99+ langs incl. CN/JP, just an HTTP call (equal for Python or Java).
faster-whisper (large-v3-turbo on GPU, INT8 small on CPU) — no quota, free forever.
Qwen3-ASR (52 langs + 22 CN dialects, ~92 ms TTFT) or VibeVoice-ASR (60-min single-pass + diarization).
SenseVoice (CN+JP, 15× faster) · Paraformer (Mandarin) · Kotoba-Whisper (JP).
Vosk (~50 MB, Java+Android+iOS) or Moonshine (<200 ms edge).
Voxtral Small (2.9% AA-WER) · IBM Granite 4.1 2B (5.33% WER) · Canary-Qwen 2.5B (5.63%).