Italian ASR

Open-source models that turn Italian audio into text, locally, on your own hardware, with optional speaker diarization. Brought to you by the LocalAI team.

🇮🇹 In italiano · leggi qui

Una famiglia di modelli open-source per trascrivere audio in italiano, da eseguire in locale sul proprio hardware, con diarizzazione (etichette parlante) opzionale. Realizzata dal team di LocalAI.

Perché esistono questi modelli. La maggior parte dei modelli Whisper italiani disponibili online sono addestrati solo su Common Voice (persone che leggono frasi al microfono di casa). Vanno benissimo nei benchmark di lettura, ma collassano sull'audio reale: riunioni, telefonate, podcast tornano pieni di ripetizioni allucinate e frasi inventate. Abbiamo addestrato una famiglia completa pensata anche per il parlato spontaneo.

Quale modello scegliere:

Scenario	Modello consigliato
🎙️ Riunioni, chiamate, podcast, audio reale	`whisper-large-v3-turbo-it-yodas-only`
📚 Audiolibri, notizie, atti parlamentari	`whisper-large-v3-it-multi`
🗣️ Servono le etichette parlante (diarizzazione)	`vibevoice-asr-it-yodas-only`

Esempio rapido (Python):

pip install faster-whisper

from faster_whisper import WhisperModel

model = WhisperModel("LocalAI-io/whisper-large-v3-turbo-it-yodas-only-ct2-int8",
                    device="cpu", compute_type="int8")  # oppure device="cuda"
segments, _ = model.transcribe("audio.mp3", language="it")
for s in segments:
    print(f"[{s.start:5.1f}s] {s.text}")

Pronti per qualunque runtime. Ogni modello è pubblicato in tutti i formati di quantizzazione che la community usa davvero, non devi convertire niente a mano:

safetensors per 🤗 Transformers (precisione piena, PyTorch).
CTranslate2 INT8 per faster-whisper / WhisperX / LocalAI, circa 4× più piccolo, ottimo su CPU.
whisper.cpp GGML (f16/q4_0/q5_0/q8_0) per whisper.cpp e LocalAI, perfetto per edge / Raspberry Pi / mobile.
GGUF per la famiglia VibeVoice via vibevoice.cpp.

Tutti i 75 repo HuggingFace sono qui: huggingface.co/LocalAI-io. Tutto sotto licenza MIT, nessuna chiave API, nessuna telemetria, gira completamente in locale.

Per la documentazione completa (tutte le varianti, i benchmark, i dettagli tecnici), continua a leggere in inglese qui sotto. 👇

Quick start

Pick the path that matches your stack. All models live on the LocalAI-io HuggingFace organization and run fully offline.

Python (faster-whisper) is recommended for most users:

pip install faster-whisper

from faster_whisper import WhisperModel

model = WhisperModel("LocalAI-io/whisper-large-v3-turbo-it-yodas-only-ct2-int8",
                    device="cpu", compute_type="int8")  # or device="cuda"
segments, _ = model.transcribe("audio.mp3", language="it")
for s in segments:
    print(f"[{s.start:5.1f}s] {s.text}")

Command line (whisper.cpp), single-binary, CPU or CUDA:

hf download LocalAI-io/whisper-large-v3-turbo-it-yodas-only-ggml \
  ggml-model-q5_0.bin --local-dir .
./whisper-cli -m ggml-model-q5_0.bin -l it -f audio.wav

LocalAI server: OpenAI-compatible HTTP endpoint. See the LocalAI section below for ready-made YAML configs and curl example.

Need speaker labels? Use the VibeVoice variant; it produces [Speaker 0/1/2…] text segments natively.

What's in this repo

A family of Italian-tuned models in two architecture families and three deployment formats:

Whisper family, every model size (tiny → large-v3-turbo) × four training recipes (CV-only, Multi, Multi+YODAS, YODAS-only). Best raw throughput on GPU.
VibeVoice ASR family, a Qwen2-based architecture with native speaker diarization for conversational audio. Best when you need speaker labels or fewer hallucinated repetitions.

Pre-quantized for every runtime

Half the reason this catalog is large is that we ship each fine-tune in every quantization format the community actually uses, so you don't have to download a PyTorch checkpoint and convert it yourself. Pick the runtime that fits your hardware:

safetensors (full precision), for 🤗 Transformers, research, and PyTorch fine-tuning.
CTranslate2 INT8 (~4× smaller, CPU-friendly), for faster-whisper / WhisperX / LocalAI's whisperx backend. Best on CPU and the default for production transcription.
whisper.cpp GGML in four bit-widths (f16, q4_0, q5_0, q8_0), for whisper.cpp, pywhispercpp, LocalAI's whisper backend, mobile / edge / Raspberry-Pi-style deployments. q5_0 is the usual sweet spot.
GGUF (q4_0, q5_0, q8_0) for vibevoice.cpp, covers the VibeVoice family on the same C++/CUDA runtime.

That's a total of 75 repos on HuggingFace, all maintained in lock-step so the q5_0 GGML matches the int8 CTranslate2 build matches the safetensors source.

Training Recipes

Why the training data matters. A speech-to-text model is only good at the kind of speech it has heard during training. A model trained on people carefully reading sentences out loud will look fantastic on benchmark numbers and then make up entire sentences when you hand it a real meeting where two humans talk over each other. So we trained four versions of each model on four different "diets" of audio, ranging from clean read speech to messy real-world conversations.

The four datasets we mix and match:

Common Voice (CV), Mozilla's crowdsourced corpus. People reading sentences on a phone or laptop at home. Clean, varied speakers, but it's read speech, not conversational.
MLS (Multilingual LibriSpeech), long-form audiobook readings. Professional narrators, no overlap, no hesitations.
VoxPopuli (VoxPopuli), European Parliament recordings. Formal speech, but real microphones in a real room.
YODAS-Granary (YODAS), YouTube clips, pseudo-labeled. Real people talking spontaneously: meetings, podcasts, vlogs. Italian as it's actually spoken.

The four mixes:

Variant	What it's trained on (plain-English)	Best for	Samples
CV-only (`*-it`)	Just Common Voice. People reading sentences at home.	CV-style benchmark WER. Falls apart on real conversations.	173k
Multi (`*-it-multi`)	Common Voice + audiobooks + parliament. All read or formal speech.	Audiobooks, news, parliamentary transcripts. Stable across read-speech domains.	255k
Multi + YODAS (`*-it-multi-yodas-asr_only-nocv`)	Audiobooks + parliament + YouTube spontaneous Italian (~87k clips).	Best when you don't know what kind of audio you'll get.	164k
YODAS-only (`*-it-yodas-only`)	YouTube spontaneous Italian only (240k clips). 100% real conversational audio.	Real-world meetings, calls, podcasts. Read-speech benchmark numbers regress (no audiobooks in training).	240k

Recommended pick:

🎙️ Meetings, calls, podcasts, real-world Italian: *-it-yodas-only (large-v3-turbo is the sweet spot for production)
📚 Audiobooks, news, parliamentary: *-it-multi
🎯 CV-style benchmarks only: *-it
🔀 Mixed/unknown audio: *-it-multi-yodas-asr_only-nocv or *-it-multi
🗣️ Multi-speaker / need diarization: vibevoice-asr-it-yodas-only (see VibeVoice section)

Models

Each link below points to a HuggingFace repo. The ggml repos contain four quantizations (f16, q4_0, q5_0, q8_0), pick by your size/quality budget. The newer YODAS-trained ggml repos ship q4_0, q5_0, q8_0 (no f16, it consistently failed to upload via the HF Xet protocol; the q8_0 is near-lossless in practice).

CV-only and Multi (all sizes)

Base	Params	CV-only (`*-it`)	Multi (`*-it-multi`)
`openai/whisper-tiny`	39M	safetensors · ct2-int8 · ggml	safetensors · ct2-int8 · ggml
`openai/whisper-base`	74M	safetensors · ct2-int8 · ggml	safetensors · ct2-int8 · ggml
`openai/whisper-small`	244M	safetensors · ct2-int8 · ggml	safetensors · ct2-int8 · ggml
`openai/whisper-medium`	769M	safetensors · ct2-int8 · ggml	safetensors · ct2-int8 · ggml
`openai/whisper-large-v3`	1.55B	safetensors · ct2-int8 · ggml	safetensors · ct2-int8 · ggml
`openai/whisper-large-v3-turbo`	809M	safetensors · ct2-int8 · ggml	safetensors · ct2-int8 · ggml

Multi + YODAS and YODAS-only (conversational coverage)

Base	Params	Multi + YODAS (`*-it-multi-yodas-asr_only-nocv`)	YODAS-only (`*-it-yodas-only`)
`openai/whisper-tiny`	39M	safetensors · ct2-int8 · ggml	safetensors · ct2-int8 · ggml
`openai/whisper-base`	74M	safetensors · ct2-int8 · ggml	safetensors · ct2-int8 · ggml
`openai/whisper-small`	244M	safetensors · ct2-int8 · ggml	safetensors · ct2-int8 · ggml
`openai/whisper-medium`	769M	safetensors · ct2-int8 · ggml	safetensors · ct2-int8 · ggml
`openai/whisper-large-v3`	1.55B	safetensors · ct2-int8 · ggml	safetensors · ct2-int8 · ggml
`openai/whisper-large-v3-turbo`	809M	safetensors · ct2-int8 · ggml	safetensors · ct2-int8 · ggml

Note: smaller *-yodas-only sizes (tiny/base/small/medium) are inherently loop-prone on conversational audio compared to large-v3/large-v3-turbo. They're shipped for completeness and edge / CPU-only deployments, for the cleanest output on real-world Italian speech, prefer large-v3-turbo-it-yodas-only.

Results

Word Error Rate (%) on the Common Voice 25.0 Italian test set (15,184 samples). Lower is better.

Base	Params	CV-only	Multi
`whisper-tiny`	39M	27.1%	29.4%
`whisper-base`	74M	19.2%	21.4%
`whisper-small`	244M	13.0%	15.6%
`whisper-medium`	769M	10.5%	12.4%
`whisper-large-v3`	1.55B	,	,
`whisper-large-v3-turbo`	809M	,	,

Large-v3 and large-v3-turbo results: these variants were trained with evaluation disabled (to avoid VRAM fragmentation during generation on very long sequences). Benchmarks will be added once we finish running the standalone eval pass, they are expected to improve on the whisper-medium numbers above.

Note on the Multi column: the *-multi models are evaluated on the CV-only test set here for an apples-to-apples comparison with the *-it variants. They were trained on a more diverse mix (CV + MLS + VoxPopuli), so CV-only WER slightly under-represents their real-world robustness on non-Common-Voice audio. If your deployment sees audiobooks, parliamentary recordings, or other non-crowdsourced speech, prefer the *-multi variant.

Training details: 10,000 steps, learning rate 1e-5, 500 warmup steps, bf16 on an NVIDIA GB10 (128GB unified memory). Common Voice 25.0 Italian was published 2026-03-09 and is the latest release at the time of training.

Real-world conversational benchmark

To complement the read-speech CV numbers above, we also benchmarked the YODAS-trained variants on an internal ~17-minute Italian conversational recording held back as a private real-world test set. The audio itself is not published, only the head-to-head numbers below. All runs on a single NVIDIA GB10 with CUDA: Whisper at q5_0 via whisper.cpp (CUDA backend), VibeVoice at q4_0 via vibevoice.cpp master (resident-KV + flash-attention).

Model	Wall time	RTF (lower is faster)	4+ word loop runs	Native diarization
`whisper-large-v3-it-yodas-only` (q5_0)	64.8 s	0.065 ⚡	7	,
`whisper-large-v3-turbo-it-yodas-only` (q5_0)	73.2 s	0.073	62	,
`vibevoice-asr-it-yodas-only` (q4_0)	619.9 s	0.622	5	✅

A 4+ word loop run counts each position where the same lowercased token appears four times in a row, a crude but consistent proxy for hallucinated repetitions on tough audio. Lower is better.

Takeaways:

For raw throughput, whisper-large-v3-it-yodas-only is unbeatable on GPU (~15× realtime).
For loop-free transcripts and free speaker diarization, vibevoice-asr-it-yodas-only wins, at the cost of being only ~1.6× realtime.
Avoid whisper-large-v3-turbo-it-yodas-only on this kind of audio, the smaller decoder degrades into long repetition runs (62 vs 7 for non-turbo), and is also slightly slower than non-turbo on GPU because of the extra fallback decode steps. Use it for read-speech where decoder length matters less.

Usage

LocalAI

LocalAI supports these models natively through the whisperx backend. Pick the variant that matches your accuracy / latency budget, drop the corresponding YAML into your LocalAI models/ directory, and transcribe via the OpenAI-compatible /v1/audio/transcriptions endpoint.

Quick hardware guide:

tiny / base, CPU-only edge devices, real-time on a Raspberry Pi.
small / medium, most servers, great accuracy / speed tradeoff.
large-v3-turbo, large-v3 quality at roughly half the latency.
large-v3, highest accuracy, when it matters most.

After setting up a model below, transcribe audio with:

curl http://localhost:8080/v1/audio/transcriptions \
  -H "Content-Type: multipart/form-data" \
  -F file="@audio.mp3" \
  -F model="<the-name-from-the-yaml-below>"

tiny (39M)

whisperx-tiny-it, Common Voice only

name: whisperx-tiny-it
backend: whisperx
known_usecases:
  - transcript
parameters:
  model: LocalAI-io/whisper-tiny-it-ct2-int8
  language: it

whisperx-tiny-it-multi, multi-dataset

name: whisperx-tiny-it-multi
backend: whisperx
known_usecases:
  - transcript
parameters:
  model: LocalAI-io/whisper-tiny-it-multi-ct2-int8
  language: it

whisperx-tiny-it-multi-yodas, multi + YODAS conversational data

name: whisperx-tiny-it-multi-yodas
backend: whisperx
known_usecases:
  - transcript
parameters:
  model: LocalAI-io/whisper-tiny-it-multi-yodas-asr_only-nocv-ct2-int8
  language: it

base (74M)

whisperx-base-it, Common Voice only

name: whisperx-base-it
backend: whisperx
known_usecases:
  - transcript
parameters:
  model: LocalAI-io/whisper-base-it-ct2-int8
  language: it

whisperx-base-it-multi, multi-dataset

name: whisperx-base-it-multi
backend: whisperx
known_usecases:
  - transcript
parameters:
  model: LocalAI-io/whisper-base-it-multi-ct2-int8
  language: it

whisperx-base-it-multi-yodas, multi + YODAS conversational data

name: whisperx-base-it-multi-yodas
backend: whisperx
known_usecases:
  - transcript
parameters:
  model: LocalAI-io/whisper-base-it-multi-yodas-asr_only-nocv-ct2-int8
  language: it

small (244M)

whisperx-small-it, Common Voice only

name: whisperx-small-it
backend: whisperx
known_usecases:
  - transcript
parameters:
  model: LocalAI-io/whisper-small-it-ct2-int8
  language: it

whisperx-small-it-multi, multi-dataset

name: whisperx-small-it-multi
backend: whisperx
known_usecases:
  - transcript
parameters:
  model: LocalAI-io/whisper-small-it-multi-ct2-int8
  language: it

whisperx-small-it-multi-yodas, multi + YODAS conversational data

name: whisperx-small-it-multi-yodas
backend: whisperx
known_usecases:
  - transcript
parameters:
  model: LocalAI-io/whisper-small-it-multi-yodas-asr_only-nocv-ct2-int8
  language: it

medium (769M)

whisperx-medium-it, Common Voice only

name: whisperx-medium-it
backend: whisperx
known_usecases:
  - transcript
parameters:
  model: LocalAI-io/whisper-medium-it-ct2-int8
  language: it

whisperx-medium-it-multi, multi-dataset

name: whisperx-medium-it-multi
backend: whisperx
known_usecases:
  - transcript
parameters:
  model: LocalAI-io/whisper-medium-it-multi-ct2-int8
  language: it

whisperx-medium-it-multi-yodas, multi + YODAS conversational data

name: whisperx-medium-it-multi-yodas
backend: whisperx
known_usecases:
  - transcript
parameters:
  model: LocalAI-io/whisper-medium-it-multi-yodas-asr_only-nocv-ct2-int8
  language: it

large-v3 (1.55B)

whisperx-large-v3-it, Common Voice only

name: whisperx-large-v3-it
backend: whisperx
known_usecases:
  - transcript
parameters:
  model: LocalAI-io/whisper-large-v3-it-ct2-int8
  language: it

whisperx-large-v3-it-multi, multi-dataset

name: whisperx-large-v3-it-multi
backend: whisperx
known_usecases:
  - transcript
parameters:
  model: LocalAI-io/whisper-large-v3-it-multi-ct2-int8
  language: it

whisperx-large-v3-it-multi-yodas, multi + YODAS conversational data

name: whisperx-large-v3-it-multi-yodas
backend: whisperx
known_usecases:
  - transcript
parameters:
  model: LocalAI-io/whisper-large-v3-it-multi-yodas-asr_only-nocv-ct2-int8
  language: it

whisperx-large-v3-it-yodas-only, pure YODAS (best on conversational)

name: whisperx-large-v3-it-yodas-only
backend: whisperx
known_usecases:
  - transcript
parameters:
  model: LocalAI-io/whisper-large-v3-it-yodas-only-ct2-int8
  language: it

large-v3-turbo (809M)

whisperx-large-v3-turbo-it, Common Voice only

name: whisperx-large-v3-turbo-it
backend: whisperx
known_usecases:
  - transcript
parameters:
  model: LocalAI-io/whisper-large-v3-turbo-it-ct2-int8
  language: it

whisperx-large-v3-turbo-it-multi, multi-dataset

name: whisperx-large-v3-turbo-it-multi
backend: whisperx
known_usecases:
  - transcript
parameters:
  model: LocalAI-io/whisper-large-v3-turbo-it-multi-ct2-int8
  language: it

whisperx-large-v3-turbo-it-yodas-only, pure YODAS (recommended for conversational)

name: whisperx-large-v3-turbo-it-yodas-only
backend: whisperx
known_usecases:
  - transcript
parameters:
  model: LocalAI-io/whisper-large-v3-turbo-it-yodas-only-ct2-int8
  language: it

Transformers

from transformers import pipeline

pipe = pipeline(
    "automatic-speech-recognition",
    model="LocalAI-io/whisper-medium-it-multi",
)
result = pipe("audio.mp3", generate_kwargs={"language": "it", "task": "transcribe"})
print(result["text"])

faster-whisper / WhisperX

The -ct2-int8 repos are drop-in replacements for the upstream CT2 Whisper models:

from faster_whisper import WhisperModel

model = WhisperModel(
    "LocalAI-io/whisper-medium-it-multi-ct2-int8",
    device="cpu",
    compute_type="int8",
)
segments, info = model.transcribe("audio.mp3", language="it")
for segment in segments:
    print(f"[{segment.start:.1f}s - {segment.end:.1f}s] {segment.text}")

whisper.cpp

Each -ggml repo contains four quantizations: ggml-model-f16.bin, ggml-model-q8_0.bin, ggml-model-q5_0.bin, ggml-model-q4_0.bin. Pick the one that fits your size / quality budget, q5_0 is the usual sweet spot, q8_0 is near-lossless, q4_0 is smallest.

huggingface-cli download LocalAI-io/whisper-medium-it-multi-ggml ggml-model-q5_0.bin --local-dir .
./whisper-cli -m ggml-model-q5_0.bin -f audio.wav -l it

These models also work with pywhispercpp and any other whisper.cpp-based runtime.

VibeVoice ASR (Italian)

In addition to the Whisper variants above, we ship a VibeVoice ASR fine-tune trained on the same YODAS-only Italian mix (LoRA, ~10k steps on the GB10). VibeVoice is a different architecture (Qwen2-based with native diarization-aware decoding), useful when you want speaker-segmented output, or fewer hallucinated loops on tough conversational audio. On our private real-world conversational test set it produced fewer 4+ word repetition runs than Whisper large-v3-yodas (5 vs 7), with native speaker diarization in the output.

Variant	Repo	Notes
Merged model	LocalAI-io/vibevoice-asr-it-yodas-only	Drop-in replacement for `microsoft/VibeVoice-ASR`.
LoRA adapter	LocalAI-io/vibevoice-asr-it-yodas-only-lora	Apply on top of the upstream base; ~tens of MB.
GGUF (vibevoice.cpp)	LocalAI-io/vibevoice-asr-it-yodas-only-ggml	`q4_0`, `q5_0`, `q8_0` quantizations for vibevoice.cpp.

Transformers

from vibevoice.modular.modeling_vibevoice_asr import VibeVoiceASRForConditionalGeneration
from vibevoice.processor.vibevoice_asr_processor import VibeVoiceASRProcessor
import torch

processor = VibeVoiceASRProcessor.from_pretrained("LocalAI-io/vibevoice-asr-it-yodas-only")
model = VibeVoiceASRForConditionalGeneration.from_pretrained(
    "LocalAI-io/vibevoice-asr-it-yodas-only",
    dtype=torch.bfloat16,
    attn_implementation="sdpa",  # or "flash_attention_2" if installed
)

vibevoice.cpp (GGUF)

Build vibevoice.cpp once (use -DVIBEVOICE_GGML_CUDA=ON for GPU), download a GGUF + the Qwen2.5 tokenizer GGUF, and run:

hf download LocalAI-io/vibevoice-asr-it-yodas-only-ggml \
  vibevoice-asr-it-yodas-only-q5_0.gguf --local-dir models/

./build/bin/vibevoice-cli asr \
  --model     models/vibevoice-asr-it-yodas-only-q5_0.gguf \
  --tokenizer models/tokenizer.gguf \
  --audio     audio.wav \
  --max-new-tokens 8000 --verbose

The output is a JSON array of diarized segments (Start, End, Speaker, Content).

Datasets

Common Voice 25.0 Italian, crowdsourced read speech, the backbone of the CV-only and Multi training mixes.
Multilingual LibriSpeech (MLS) Italian, audiobook readings, adds long-form narration to the Multi mix.
VoxPopuli Italian, European Parliament recordings, adds parliamentary speech to the Multi mix.
YODAS-Granary Italian, YouTube-sourced spontaneous speech, pseudo-labeled with faster-whisper-large-v3 and cleaned with Qwen2.5-7B-Instruct. Provides the conversational coverage that the Multi mix alone is missing. Used in two slices: asr_only (~87k clips) for the Multi+YODAS variants, and asr_only + ast (200k cap) for the YODAS-only variants.

License

Models and code are released under the MIT license, consistent with the upstream OpenAI Whisper release.

Citation

If you use these models in academic work or want to cite this release, please use the following BibTeX entry:

@software{italian_asr_2026,
  author    = {Di Giacinto, Ettore and Palethorpe, Richard},
  title     = {{Italian ASR}: An open family of Italian speech-to-text models
               (Whisper and VibeVoice fine-tunes) with a C++ inference engine
               for VibeVoice},
  year      = {2026},
  publisher = {LocalAI},
  url       = {https://github.com/mudler/italian-asr},
  note      = {Models on HuggingFace at \url{https://huggingface.co/LocalAI-io};
               vibevoice.cpp engine at \url{https://github.com/mudler/vibevoice.cpp}}
}

Plain-text:

Di Giacinto, E. and Palethorpe, R. (2026). Italian ASR: An open family of Italian speech-to-text models (Whisper and VibeVoice fine-tunes) with a C++ inference engine for VibeVoice. LocalAI. https://github.com/mudler/italian-asr

Acknowledgements

Built by Ettore Di Giacinto and Richard Palethorpe of the LocalAI team. If you find these models useful, please star LocalAI on GitHub, it genuinely helps the project.

Name		Name	Last commit message	Last commit date
Latest commit History 1 Commit
assets		assets
README.md		README.md

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Repository files navigation

Italian ASR

Quick start

What's in this repo

Pre-quantized for every runtime

Training Recipes

Models

CV-only and Multi (all sizes)

Multi + YODAS and YODAS-only (conversational coverage)

Results

Real-world conversational benchmark

Usage

LocalAI

tiny (39M)

base (74M)

small (244M)

medium (769M)

large-v3 (1.55B)

large-v3-turbo (809M)

Transformers

faster-whisper / WhisperX

whisper.cpp

VibeVoice ASR (Italian)

Transformers

vibevoice.cpp (GGUF)

Datasets

License

Citation

Acknowledgements

About

Uh oh!

Releases

Packages

Uh oh!

Contributors

Uh oh!

Uh oh!

Folders and files

Latest commit

History

Repository files navigation

Italian ASR

Quick start

What's in this repo

Pre-quantized for every runtime

Training Recipes

Models

CV-only and Multi (all sizes)

Multi + YODAS and YODAS-only (conversational coverage)

Results

Real-world conversational benchmark

Usage

LocalAI

tiny (39M)

base (74M)

small (244M)

medium (769M)

large-v3 (1.55B)

large-v3-turbo (809M)

Transformers

faster-whisper / WhisperX

whisper.cpp

VibeVoice ASR (Italian)

Transformers

vibevoice.cpp (GGUF)

Datasets

License

Citation

Acknowledgements

About

Topics

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Packages