Datasetpublic

Gstam21/live-translator-packs

Live Translator — Offline Model Packs On-demand offline model packs for the Live Translator app. Downloaded per language pair; the shared common/ bundle installs once on the first pair. Layout common/ # downloaded ONCE, on the first pair asr/ # Whisper tiny int8 (multilingual ASR + language id), sherpa-onnx whisper-encoder.onnx whisper-decoder.onnx whisper-tokens.txt vad/silero_vad.onnx # Silero… See the full description on the dataset page: https://huggingface.co/datasets/Gstam21/live-translator-packs.

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes73downloads

Dataset Card

Live Translator — Offline Model Packs

On-demand offline model packs for the Live Translator app. Downloaded per language pair; the shared common/ bundle installs once on the first pair.

Layout

common/                       # downloaded ONCE, on the first pair
  asr/                        # Whisper tiny int8 (multilingual ASR + language id), sherpa-onnx
    whisper-encoder.onnx
    whisper-decoder.onnx
    whisper-tokens.txt
  vad/silero_vad.onnx         # Silero VAD
  denoiser/gtcrn.onnx         # GTCRN speech denoiser
  espeak-ng-data.tar          # shared phonemizer data (untar to common/espeak-ng-data/)
  tts-kokoro/                 # Kokoro-82M int8 — shared multilingual voice
    model.onnx                #   covers en, es, fr, it, pt, hi, ja, zh
    voices.bin
    tokens.txt
    lexicon-us-en.txt

<pair>/                       # e.g. en-el/, en-es/  — downloaded per pair
  nmt-<a>-<b>/                # CTranslate2 int8 NMT, one dir per direction
    model.bin
    config.json
    shared_vocabulary.json
    source.spm                # SentencePiece tokenizers (Marian/OPUS-MT)
    target.spm
  nmt-<b>-<a>/
  tts-<lang>/                 # ONLY for languages Kokoro does NOT cover (e.g. Greek → Piper)
    model.onnx                #   en/es come from common/tts-kokoro, so en-es has no tts dir
    tokens.txt

Pipeline / runtimes

—ASR: Whisper tiny int8 via sherpa-onnx (onnxruntime).
—NMT: CTranslate2 int8 (OPUS-MT / Marian). Requires the CT2 runtime on device (dart:ffi → libctranslate2). Tokenize with the .spm SentencePiece models; decode with beam search (beam=5, no-repeat-ngram, length penalty). Feed one sentence per call — Marian emits a single sentence per translate.
—TTS: Kokoro int8 (sherpa-onnx) for en/es/…; Piper medium for Greek.

Sizes

Item	Size
common (once)	~284 MB (ASR 98 · Kokoro 166 · espeak 17 · vad/denoise 1)
en-el	~196 MB (CT2 NMT 140 · Piper-el-medium 60)
en-es	~158 MB (CT2 NMT only; en+es voices from common Kokoro)
First download (common + en-el)	~480 MB
+ en-es	+158 MB

Source models

Pack	Source
nmt-en-el	`Helsinki-NLP/opus-mt-en-el` → CT2 int8
nmt-el-en	`Helsinki-NLP/opus-mt-grk-en` → CT2 int8
nmt-en-es	`Helsinki-NLP/opus-mt-en-es` → CT2 int8
nmt-es-en	`Helsinki-NLP/opus-mt-es-en` → CT2 int8
asr	Whisper tiny int8 (sherpa-onnx)
tts-kokoro	`csukuangfj/kokoro-int8-multi-lang-v1_1`
tts-el	`rhasspy/piper-voices` el_GR-rapunzelina-medium

Served at https://huggingface.co/datasets/Gstam21/live-translator-packs/resolve/main/<path>.