Skip to content

models.md

Latest commit

 

History

History
131 lines (99 loc) · 6 KB

models.md

File metadata and controls

131 lines (99 loc) · 6 KB
Raw

Model selection

Seda exposes exact upstream model IDs. A variant selects a concrete runtime artifact; an immutable revision and runtime complete the resolved identity. Profiles remain optional aliases, not the primary API.

Runtime Model ID Variant Alias Download Language mode Output
Browser onnx-community/moonshine-tiny-ONNX q4 WebGPU / q8 WASM — ~55 MB fixed English buffered revisions, punctuation
Native nvidia/parakeet_realtime_eou_120m-v1 q4_k compact 129,133,984 bytes fixed English true streaming, EOU, word times
Native nvidia/nemotron-3.5-asr-streaming-0.6b q4_k balanced 718,102,624 bytes prompted, 32 locales, auto true streaming, punctuation, word times
Native nvidia/nemotron-3.5-asr-streaming-0.6b q8_0 quality 983,696,512 bytes prompted, 32 locales, auto same model, higher weight precision

Prepare and start by ID:

seda prepare \
  --model-id nvidia/nemotron-3.5-asr-streaming-0.6b \
  --variant q4_k
seda serve \
  --model-id nvidia/nemotron-3.5-asr-streaming-0.6b \
  --variant q4_k

Then select de-DE, ja-JP, or auto on each transcription or live session. Changing the language prompt does not prepare or reload the model.

Language modes

Seda reports language behavior instead of assuming every multilingual model works the same way:

Mode Meaning When language is chosen
fixed One model recognizes one language Fixed by the model
prompted One resident model accepts a language prompt Every call or session
automatic The runtime detects language No caller prompt
checkpoint A separate checkpoint exists per language When choosing/preparing the model ID

Applications should build their picker from capabilities().language, not from hard-coded product tiers.

Browser: Moonshine Tiny

@bearlyai/seda-browser pins onnx-community/moonshine-tiny-ONNX at revision a6da1241cd305dcd64eab1edbd615f2bb9aabb95. The English Moonshine model and ONNX distribution are MIT licensed. Transformers.js 3.7.6 is Apache-2.0.

The Worker uses the full-precision encoder and Q8 merged decoder on WASM. WebGPU uses a Q4 merged decoder. WebGPU is preferred automatically and WASM is the portable fallback. The model is cached by the browser rather than placed in Seda’s native data directory.

Moonshine Tiny is the browser default because its roughly 55 MB working set is materially more deployable than the native Parakeet catalog and it has a proven Transformers.js browser path. It is not cache-aware streaming: Seda periodically decodes the current utterance, exposes those outputs as replaceable buffered revisions, and performs a final decode on commit. One utterance is limited to the model’s documented 30 seconds.

The compact native model is attractive for push-to-talk because NVIDIA describes it as a 120M-parameter cache-aware streaming recognizer with 80–160 ms latency and explicit end-of-utterance output. It is English-only and intentionally has no punctuation or capitalization: https://huggingface.co/nvidia/parakeet_realtime_eou_120m-v1

Nemotron 3.5 is the multilingual default because it is cache-aware true streaming, supports automatic language detection, and includes punctuation and capitalization. Seda advertises only the 19 transcription-ready and 13 broad-coverage locales, not the eight adaptation-only tokenizer locales: https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b

The GGUF files are community quantizations from: https://huggingface.co/mudler/parakeet-cpp-gguf

The pinned native runtime is parakeet.cpp v0.4.0: https://github.com/mudler/parakeet.cpp/tree/v0.4.0

Why Parakeet remains native

Parakeet gives Seda a useful launch combination: a small English streaming model, a genuinely streaming multilingual model, word timestamps, a compact C ABI, and prebuilt libraries for the three initial OS targets. More adapters would increase surface area before the protocol and product behavior are proven.

Alternatives evaluated

sherpa-onnx

sherpa-onnx has the broadest deployment matrix: desktop, mobile, Node, embedded, and WebAssembly, with both streaming and non-streaming recognizers: https://github.com/k2-fsa/sherpa-onnx

Why not v0.1: it is a toolkit and model ecosystem rather than one curated model experience. It is a good future adapter for WASM, ARM, mobile, and languages not covered by the launch catalog.

whisper.cpp

whisper.cpp is mature, portable, multilingual, and an excellent batch fallback: https://github.com/ggml-org/whisper.cpp

Why not the default: its common “streaming” integrations repeatedly decode a rolling buffer. That can be useful, but Seda reports it as buffered streaming rather than presenting it as the same behavior as a cache-aware streaming recognizer.

Browser constraints

“Runs in a browser” does not mean every device has equal latency or memory:

  • Cache API storage is persistent but browser eviction can require a reinstall.
  • WebGPU availability and performance vary; WASM is the compatibility tier.
  • Low-memory mobile browsers may fail the model allocation.
  • The 30-second limit is enforced before memory can grow without bound.
  • English is the only advertised browser language in v0.2.

The GitHub model lane downloads the pinned revision and recognizes a checksum-verified real speech fixture through Chromium WASM. The regular browser lane verifies the session contract in Chromium, Firefox, and WebKit.

Licensing

Seda code and Transformers.js are Apache-2.0. English Moonshine and its ONNX distribution are MIT. parakeet.cpp is MIT. The GGUF distribution identifies itself as CC-BY-4.0. The compact base model uses NVIDIA’s Open Model License; Nemotron 3.5 uses OpenMDW 1.1. Applications distributing or auto-downloading weights must evaluate and retain the applicable upstream notices. Seda records native licenses in models/catalog.json and browser pins in packages/browser/src/models.ts, but does not provide legal advice.