Hermes Intelligence · Trending-OSS sweep · decision-grade

Trending Open-Source ASR — Meeting-Pipeline Fit

2026-07-08 · extends the 2026-07-07 engine refresh · routes to metis #110 Track D
Swept ~120 repos (3 topic queries + recent-created-by-stars) → 2 verify lanes (models · frameworks/velocity)

Lane E — new models fitLane F — velocity + frameworks + TTS

One-line verdict. The trend is real and active — but it doesn't change the answer for our constraints. Icelandic is the moat. Every trending new model was checked against its enumerated language list; none includes Icelandic. Keep faster-whisper-large-v3 in skynet.

TL;DR

Nothing new beats whisper-v3 for bilingual IS/EN streaming self-host. Qwen3-ASR, FunASR/SenseVoice, SoulX-Transcriber — all verified absent Icelandic (enumerated lists, not "multilingual" hand-waving).
The one trending model WITH Icelandic — Meta Omnilingual ASR (Apache-2.0, CER 3.8% on IS) — is batch-only + stale since Dec 2025. Not a drop-in.
Nothing beats skynet as a server either. Trending frameworks are wrappers/desktop apps (wrong shape) or generic faster-whisper servers with no roster/meeting glue (WhisperLive = closest, but you'd rebuild skynet's integration).
Velocity, measured honestly: only Meetily (+389★/wk) and voicebox (+378★/wk) are provably surging. The irony: the hottest meeting-transcription project is Meetily — the tool you're leaving. Star velocity ≠ fit.
TTS (you said "audio generation") — adjacent, not the transcription layer. Zonos2 (MIT, no IS), MisoTTS (English-only), dots.tts (IS unknown). None confirmed IS.
Trend to watch: Parakeet-class streaming transducers — but Parakeet is IS-blind, so it doesn't help us either.

New models — the Icelandic wall (Lane E, primary-source verified)

Model	IS?	Stream	License	Verdict
Qwen3-ASR (0.6B/1.7B)	N — 30 langs+22 dialects enumerated, no IS	Y	Apache-2.0	DISQUALIFIED on IS — best-fit otherwise, IS the sole blocker
FunASR / SenseVoice (19k★)	N — "50+" is a union across sub-models; none has IS	Y (server + OpenAI API + diarization)	MIT	DISQUALIFIED on IS — best skynet replacement if any model had IS
SoulX-Transcriber (joint speaker+ASR)	N — ZH/EN only	N — batch	Apache-2.0	DISQUALIFIED — no IS, no streaming, 30B footprint
Meta Omnilingual ASR (discovered)	Y — isl_Latn, CER 3.8%, confirmed	N — batch-only	Apache-2.0	VIABLE only if streaming dropped — stale since Dec 2025

Why they all miss IS: the 2026 wave optimizes for the big commercial-language set. Icelandic (~370k speakers) isn't in any published training mix — same pattern already confirmed for Canary/Parakeet/Nemotron/Voxtral. Structural, not a gap a newer release is likely to close soon.

Frameworks — nothing beats skynet as a server (Lane F)

skynet's streaming_whisper already is a meeting-shaped, roster-aware, multi-tenant server (per-participant speaker_id|lang framing, meeting-scoped JWT, faster-whisper, /summarize+/action-items). The trending field by shape:

Wrong shape — desktop/CLI apps: Meetily, voicebox, transcria (batch portal), livecaption (macOS/MLX).
Wrong shape — libraries: RealtimeSTT (cooling), parakeet-rs, whisper_streaming (self-declared superseded), SimulStreaming.
Right shape, no meeting integration: VoiceStreamAI (dead since Oct 2024); WhisperLive (healthy alt server — but no Jitsi/roster/JWT/summarize → rebuild skynet's glue); sherpa-onnx (Kaldi/ONNX = re-base the stack).

Velocity — what's actually surging (and why it doesn't matter here)

Project	Surge evidence	Fit
Meetily	+389★/week (Trendshift wk27, #12 overall)	❌ desktop app — the tool we're leaving
voicebox	+378★/week (wk26); 38.6k★ in 5.5mo	❌ desktop voice-I/O app
everything else	flat / cooling / too-new / unverified	—

Lesson: the star-velocity leaderboard and the pipeline-fit shortlist barely intersect. What's hot is local desktop meeting-note apps — the category you've outgrown by going server-side + roster-aware. Method note: GitHub's stargazer endpoint 403'd this session; velocity corroborated via Trendshift dated weekly pages + commit cadence — nothing called "surging" on total-star-count alone.

TTS / audio-generation — adjacent, not the transcription layer

Project	License	Icelandic?
Zyphra/ZONOS2 (MoE TTS)	MIT	lists Nordic langs — but NOT Icelandic
MisoTTS (8B emotive)	"modified MIT" / NOASSERTION	English-only; cooling
dots.tts (2B AR)	Apache-2.0	unknown / not found

Text→speech, opposite direction of the pipeline — no bearing on the meeting-transcription server. If Icelandic TTS is ever needed (e.g. a Viska voice UI), that's a separate arc; none of these cover IS today.

Recommendation

Keep faster-whisper-large-v3 in skynet. The sweep — run precisely to catch anything better — confirms the incumbent. Yesterday's bilingual plan stands: per-participant language routing + the in-meeting toggle (D7).

Two forward paths, both net-new engineering (neither clears the bar to displace working whisper-v3 today):

Fine-tune/port an IS checkpoint into FunASR's streaming server — better serving architecture (OpenAI API + diarization + MIT) while keeping IS. Revisit if skynet's serving limits ever bite.
Streaming wrapper around Meta Omnilingual's IS-capable 300M/1B tier — the only new model with confirmed Icelandic, if its CER 3.8% beats whisper-v3 on IS (head-to-head not yet run).

Watch, don't adopt: Parakeet-class streaming transducers are the model-layer momentum — but IS-blind, so not for us until someone trains Icelandic into one.

Hermes research · research/asr-trending-2026-07-08/ · lanes E + F + raw topic digests
All sources fetched 2026-07-08. Not re-verified: whisper-v3's own current IS CER (incumbent, out of scope). Star velocity via Trendshift (GitHub stargazer API access-blocked — documented in lane F).