Modelpublic

eugenehp/rlx-asr

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes13downloads

Model Card

RLX streaming Conformer ASR

Single-file RLX streaming ASR pack (frontend, VAD, Conformer, CTC, AED) as model.rlxp.

Field	Value
Hub id	`eugenehp/rlx-asr`
Kind	RLX-native weight bundle (graphs + sidecars ready for `rlx-*` crates).
RLX crate	`rlx-asr`

Quick start

bash

just fetch-rlx-asr   # or: hf download eugenehp/rlx-asr model.rlxp --local-dir weights/asr
just fetch-rlx-asr && cargo run -p rlx-asr --release -- transcribe --wav clip.wav

Primary files (use these)

—model.rlxp — 591.6 MiB

Hub ships model.rlxp only. Pack locally with just asr-pack-rlxp (can convert from a local model.gguf).

Pack layout (`.rlxp`)

Tensor pack for streaming Conformer ASR: named weights plus CTC unit list and etiquette metadata as sidecars. No ONNX graphs — the RLX graph is built at runtime.

Official RLX package format (RLXPFLAT, container v2).

text

[0..8)   magic          RLXPFLAT
[8..12)  version        u32 LE (= 2)
[12..16) flags          u32 LE (hybrid hot/warm/cold)
[16..24) toc_len        u64 LE
[24..)   TOC            JSON table of contents
         data region    64-byte aligned payloads

The TOC lists tensors (named weight blobs) and/or sidecars (files: ONNX, tokenizers, manifests, …). Sidecars are usually cold + zstd; model weights in tensor packs are hot + uncompressed for mmap. Runtime crates open the pack directly (or materialize sidecars to a temp dir for asset-only packs).

This pack

Field	Value
File	`model.rlxp` (591.6 MiB)
Manifest name	`rlx-asr`
Producer	`rlx-asr`
Container	RLXPFLAT v2, flags=0x1
Tensors	272
Sidecars	2

Tensors (hot weight region)

Weights come from the former model.gguf (rlx-asr layout). Most are f32; a few codebook / index tensors are i8.

All tensors are mmap'd from the hot region (schemes: f32×263, i8×9).

Prefix	Tensors	Stored	Role
`encoder.*`	239	540.9 MiB	Conformer encoder stack
`decoder.*`	10	43.8 MiB	AED / CTC projection (`effective_We`, …)
`ls.*`	10	5.1 MiB	Layer-scale / auxiliary layer tensors
`codebook.*`	8	1.0 MiB	Discrete codebook tables
`tp.*`	4	742.3 KiB	Token / projection helpers
`silence_fbank.*`	1	320 B	Silence filterbank reference

Sidecars (file assets)

Text metadata only — no neural graphs in sidecars.

Paths below are logical ids inside the pack (__flat__/sidecar/<id>). Cold sidecars are zstd-compressed; sizes show raw → stored.

Sidecar	Raw	Stored	Role
`units.txt`	41.7 KiB	19.5 KiB	CTC / unit vocabulary (one symbol per line)
`etiquette.json`	8.4 KiB	1.3 KiB	Pack etiquette / runtime metadata

Architecture

Pipeline: frontend/encoder → Conformer → CTC/AED decode.

Prefix	Role	dtype
`encoder.*`	Conformer stack	mostly f32
`decoder.*`	AED / CTC head	f32
`codebook.*`	discrete tables	f32 / i8
`ls.` / `tp.`	layer-scale / projections	f32

Logical tree

text

model.rlxp
├── tensors/          # hot mmap region
│   ├── encoder.*     # Conformer
│   ├── decoder.*     # AED head
│   ├── ls.* / codebook.* / tp.*
│   └── silence_fbank
└── sidecars/         # cold zstd
    ├── units.txt
    └── etiquette.json

How it is packed

just asr-pack-rlxp → rlx-asr-pack-gguf --rlxp. Prefers converting an existing local model.gguf; otherwise packs from loose sources under .cache/asr / weights/asr.

Note

Hub ships .rlxp only. A local legacy model.gguf still loads if present.

Run with RLX

Clone rlx-models, place this repo under weights/asr (or pass the path explicitly), then: