Trending Model:#1Inklingthinkingmachines⬇16kTrending Model:#2Ternary-Bonsai-27B-ggufprism-ml⬇432kTrending Model:#3Unlimited-OCRbaidu⬇2237kTrending Model:#4Bonsai-27B-ggufprism-ml⬇1405kTrending Model:#5GLM-5.2zai-org⬇545kTrending Model:#6Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUFDavidAU⬇63kTrending Model:#7Qwythos-9B-Claude-Mythos-5-1M-GGUFempero-ai⬇2133kTrending Model:#8krea2-identity-editconradlocke⬇0kTrending Model:#9Qwen3.6-35B-A3B-Uncensored-HauhauCS-AggressiveHauhauCS⬇1998kTrending Model:#10OvisOCR2ATH-MaaS⬇17kTrending Model:#1Inklingthinkingmachines⬇16kTrending Model:#2Ternary-Bonsai-27B-ggufprism-ml⬇432kTrending Model:#3Unlimited-OCRbaidu⬇2237kTrending Model:#4Bonsai-27B-ggufprism-ml⬇1405kTrending Model:#5GLM-5.2zai-org⬇545kTrending Model:#6Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUFDavidAU⬇63kTrending Model:#7Qwythos-9B-Claude-Mythos-5-1M-GGUFempero-ai⬇2133kTrending Model:#8krea2-identity-editconradlocke⬇0kTrending Model:#9Qwen3.6-35B-A3B-Uncensored-HauhauCS-AggressiveHauhauCS⬇1998kTrending Model:#10OvisOCR2ATH-MaaS⬇17k

This Realtime-Multilingual-Asr-Router Barters Tiny AI Models For Crystal Clear Multilingual Transcripts

By vramkickedin | | 2 min read
Glowing audio waveform splitting into eleven distinct colored threads.

The Realtime-Multilingual-Asr-Router is a lightweight coordinator that routes live audio between small, specialized speech recognition models instead of relying on a single large multilingual system. It detects speech boundaries with Silero VAD, identifies the spoken language using SpeechBrain, and automatically swaps transcription models whenever a language switch is detected. A rollback engine then rewinds the transcript to the switch point and re-transcribes the buffered audio so the final output is clean.

Gladia open-sourced this tool to make real-time multilingual transcription practical on modest hardware. The project tackles the fact that large multilingual ASR models are often too heavy for local machines or suffer from poor accuracy when languages switch mid-conversation. By orchestrating several tiny Zipformer models (about 100 million parameters each), the router delivers fast, CPU-friendly performance without sending audio to the cloud.

Local-first routing across 11 languages

Key Features
  • Routes audio between small monolingual Zipformer models.
  • Silero VAD detects precise speech boundaries.
  • SpeechBrain identifies the language in real time.
  • Rollback engine corrects mid-stream language switches.
  • Supports 11 languages including English, Spanish, German.
  • Lightweight design runs comfortably on CPU.
  • WebSocket and REST API for custom integrations.
  • Browser frontend for instant testing.

This release fits well for privacy-conscious professionals, prosumer GPU owners, and small agencies that need multilingual transcription under their own control. Because the system runs locally, no audio leaves the machine, making it suitable for sensitive voice data. It also provides cloud-competitive accuracy at a fraction of the resource cost, serving teams that want to avoid recurring API fees.

What developers should know

In synthetic inter-utterance code-switching tests, the router scored a 13% word error rate, beating Deepgram, ElevenLabs, and Voxtral Mini in those benchmarks. The main limitation is mid-sentence language mixing (intra-utterance switching), where accuracy drops to around 41%—still an improvement over similarly sized open-source alternatives. The MIT-licensed project offers a quick setup with a simple download script and automatic model discovery, so you can start transcribing locally in minutes.

“Rather than relying on one massive multilingual model, the system routes audio between smaller, specialized monolingual models (~100M parameters each).” — Source: Reddit

Project Links