Modelpublic

luizhrocha/parakeet-tdt-0.6b-v3-ptBR-gguf

sourceHugging Facecc-by-4.0updated 3mo agoView on Hugging Face

0likes28downloads

Model Card

parakeet-tdt-0.6b-v3-ptBR — GGUF (for parakeet.cpp)

GGUF conversion of a Brazilian-Portuguese fine-tune of NVIDIA's parakeet-tdt-0.6b-v3, for CPU inference with parakeet.cpp.

Converted with parakeet.cpp's own scripts/convert_parakeet_to_gguf.py. No weights were retrained or altered — this is a format conversion only.

file	quant	size
`ptbr-plus-q8_0.gguf`	`q8_0`	~898 MB

Why this exists

The stock parakeet-tdt-0.6b-v3 is multilingual on paper, but it takes no language prompt — you cannot tell it what language it is hearing. It auto-detects, and its detection is unreliable: on Portuguese audio it frequently emits fluent, confident English. This is a known, unresolved upstream issue (NVIDIA-NeMo/Speech#14799).

It is also span-dependent, which makes it especially treacherous. Measured on one 90-second pt-BR recording and a 25-second slice of that same audio:

tdt-v3, full 90s  → "Marco 1. Eu começo a falar imediatamente, sem nenhuma pausa…"   ✅ pt
tdt-v3, 25s slice → "Marco 2, esta is my voice and volume normal…"                    ❌ en

A pt-BR fine-tune cannot drift, because Portuguese is all it knows. The language is correct by construction rather than forced — and there is no flag to forget to set.

Measurements

On a 90-second read pt-BR script (WER against the script; number words folded to digits so "dois mil e vinte e seis" and "2026" compare equal):

model	WER	pt-BR reliability	speed (CPU)	punctuation
this model	12.7%	deterministic — cannot drift	~9× realtime	yes
`tdt-0.6b-v3`	13.3%	lottery — English on a short span	~9× realtime	yes
`nemotron-3.5-asr-streaming-0.6b` (forced `pt`)	20.7%	deterministic	~8× realtime	partial
`whisper small`	29.3%	deterministic	much slower on CPU	yes

Single speaker, single recording — treat as indicative, not a benchmark.

Usage

bash

parakeet-cli --model ptbr-plus-q8_0.gguf --decoder tdt audio.wav

No --lang needed (and none is honoured — TDT models take no prompt).

Provenance and licensing

nvidia/parakeet-tdt-0.6b-v3                     CC-BY-4.0
  └─ alexandreacff/parakeet-tdt-0.6b-v3-ptBR-plus   (fine-tune, pt-BR + podcasts,
     │                                                punctuated + synthetic)
     └─ this repo                                   (GGUF format conversion only)

—Base model: `nvidia/parakeet-tdt-0.6b-v3` — CC-BY-4.0.
—Fine-tune: `alexandreacff/parakeet-tdt-0.6b-v3-ptBR-plus` — all credit for the pt-BR adaptation goes to @alexandreacff.

Licensing note, stated plainly: the fine-tune does not declare a license of its own. It is a derivative of a CC-BY-4.0 model, and this conversion is published under CC-BY-4.0 on that basis, with attribution to both authors above. If either author objects, open a discussion here and this repo will be taken down.

Converted for use in chimr, which needs correct Portuguese captions on CPU-only machines.