niksedk/parakeet-ctc-0.6b-vi-GGUF

Parakeet CTC 0.6B Vietnamese — GGUF

GGUF / ggml conversions of `nvidia/parakeet-ctc-0.6b-Vietnamese` for the crispasr CLI from CrispStrobe/CrispASR, as used by the Crisp ASR Parakeet engine in Subtitle Edit.

• —Vietnamese only, trained on 2,000+ hours of Vietnamese speech; output has punctuation and casing.
• —FastConformer-CTC: 24-layer FastConformer encoder (d_model=1024, 8 heads, FFN 4096) + a single CTC head over a 1024-piece SentencePiece vocabulary (+1 blank). No RNN-T/TDT prediction network.
• —16 kHz mono, 80 mel bins, 8× temporal subsampling.
• —CC-BY-4.0 (upstream licence).

Converted from parakeet-ctc-0.6b-vi.nemo (sha256 b1df1b01d9e833ca15930e29d0f2834a1668701dd40b11b573b04c3cfcca3a42) with CrispASR's models/convert-stt-fastconformer-ctc-to-gguf.py, quantised with crispasr-quantize from the v0.8.31 release.

Files

All four gave the same or near-identical transcripts on FLEURS vi_vn validation clips with crispasr 0.8.31 (Apple M4, Metal).

Usage

This is a CTC model. Let crispasr auto-detect the backend, or pass --backend fastconformer-ctc. Do not pass --backend parakeet on releases before v0.8.31 — that is the transducer runtime and rejects CTC models (0.8.31 reroutes automatically).

The model punctuates by itself, so turn off crispasr's automatic FireRedPunc punctuation restoration for this backend with --punc-model none; FireRedPunc's Chinese/English vocabulary otherwise glues Vietnamese words together.

crispasr --backend fastconformer-ctc -l vi -m parakeet-ctc-0.6b-vi-q8_0.gguf --punc-model none -f audio16k.wav --output-srt

The tokenizer has space-prefixed ▁, / ▁. pieces that the model prefers, so raw output reads gần xe , và ... dàng .; strip the space before punctuation afterwards (Subtitle Edit does this automatically).