Modelpublic

AladdinChen/trans-local-speech-models

sourceHugging Faceotherupdated 3d agoView on Hugging Face

0likes

Model Card

Trans local speech models

Versioned, data-only runtime packages for local speech recognition in Trans. These are conversions of the credited upstream models, not models trained by AladdinChen. The original authors retain their rights. No endorsement is implied.

Each package is downloaded separately. Recognition runs on the user's device; this repository contains no hosted inference service, recordings, training data, app credentials, or application code.

Language	Package	Approximate download	Published upstream license
Filipino / Tagalog	FastConformer CTC INT8	175 MB	CC BY 4.0
Turkish	Whisper Small Q8_0	264 MB	CC0 1.0; underlying Whisper MIT
Finnish	Whisper Medium Q8_0	823 MB	Apache 2.0; underlying Whisper MIT
Danish	Whisper Small Q8_0	264 MB	MIT
Swedish	KB Whisper Small Q8_0	264 MB	Apache 2.0; underlying Whisper MIT
Hebrew	GigaAM Hebrew CTC INT8	320.4 MB	MIT
Greek	Whisper Turbo Q8_0	874 MB	MIT
Norwegian Bokmål	NB Whisper Small Q8_0	264 MB	Apache 2.0; underlying Whisper MIT
Serbian	Whisper Small Q8_0	264 MB	Apache 2.0; underlying Whisper MIT
Khmer	Whisper Small Q8_0	264 MB	Apache 2.0; underlying Whisper MIT
Burmese	Whisper Small Q8_0	264 MB	Apache 2.0; underlying Whisper MIT
Romanian	Whisper Small Q8_0	264 MB	Apache 2.0; underlying Whisper MIT

See LICENSES.md for source authors, licenses and modification notices. Full license texts and preserved upstream notices are in licenses/ and provenance/. Redistributors must retain the notices and meet the applicable per-model license conditions. This is not a blanket license over all files.

Packages and integrity

manifest.json records the exact upstream revisions, modification method, file sizes and SHA-256 digests. Model weights live under models/<language>/. Production clients should pin a repository commit, not the moving main branch, and verify both size and SHA-256 after download.

The Whisper packages use whisper.cpp's GGML Q8_0 format. Select the listed language explicitly and use transcription, not translation, mode. They are language-specialized fine-tunes of a multilingual architecture; conversion did not remove other languages or shrink the network. The Filipino package uses the CTC head, with dynamic INT8 MatMul/Gemm weights and sherpa-onnx metadata. Its expected input is 16 kHz mono audio.

Limitations

These packages can misrecognize names, accents, numbers, background speech and noise, or hallucinate words. Quantization and streaming adapters can change accuracy relative to publisher benchmarks. No publisher benchmark is claimed as a measurement of these converted files. Large models need device-specific memory, latency and battery testing. Norwegian Nynorsk has not been validated.

Training-data rights and deployment requirements are separate from the model weight license. Review the preserved upstream model cards for limitations. Malay, the older unlicensed Serbian Tiny checkpoint, and Mongolian remain excluded pending resolution of their licensing or provenance questions. The separately licensed Serbian Small Combined checkpoint listed above is included.

Additional packages

Language	Package	Download	Publisher-stated license
Vietnamese	Gipformer 1.5 INT8	73.2 MB	MIT
Thai	Thai Zipformer INT8	160.9 MB	Apache 2.0
Arabic	NVIDIA FastConformer Arabic INT8	131.7 MB	CC BY 4.0
Hindi	IndicConformer Hindi INT8	197.7 MB	MIT source; Apache 2.0 export
Indonesian	Indonesian Zipformer2 INT8	70.9 MB	MIT
Persian	Shenava Koochik v1.5 RNNT INT8	136.7 MB	Apache 2.0 fine-tune; NVIDIA base CC BY 4.0
Croatian	Croatian FastConformer CTC INT8	131.3 MB	CC BY 4.0
Dutch / Polish / Ukrainian / Czech	Parakeet TDT 0.6B v3 INT8	670.5 MB	CC BY 4.0
Russian	GigaAM v3 E2E RNN-T INT8	226.4 MB	MIT

Parakeet is uploaded once and shared across its listed app languages. Similarly sized Whisper packages above have different publisher-trained weights; they are not renamed copies of one checkpoint. No additional fine-tuning was done by Trans. Russian GigaAM is prepared for an on-demand app download. Adding a package here does not automatically add a language to an already installed app: the current app uses a built-in catalog and pinned download revisions.

Hebrew replacement — GigaAM-He

The current Hebrew package is GigaAM Hebrew CTC INT8, adapted from asfberlin/fast-hebrew-asr, not the Russian GigaAM checkpoint. It is smaller and faster in local testing but less accurate than Ivrit Whisper. See the linked provenance for the measured accuracy tradeoff and audio-frontend compatibility limitation. It does not supply punctuation. The previous models/he/he-whisper-q8.bin and its attribution remain available for older pinned app versions; no old model was deleted.

Four additional language-specialized Whisper packages

Serbian, Khmer, Burmese and Romanian use separate, versioned Q8_0 packages. They are not newly trained by AladdinChen. Khmer and Burmese require particular accuracy validation before production use. See each package notice for provenance and limitations. Existing models and pinned URLs have not been deleted or overwritten.