Downloads · 30 days
11
19% of all-time downloads
surus-ai/audio8-asr-0.1b-es-rioplatense
audio8-asr-0.1b-es-rioplatense is a automatic speech recognition model from surus-ai. Use it when you need speech turned into text. It is set up for transformers. The card lists the license as cc-by-nc-4.0.
Fine-tune de Audio8/Audio8-ASR-0.1B especializado en español rioplatense / latinoamericano mediante SFT (supervised fine-tuning) completo.
Downloads · 30 days
11
19% of all-time downloads
All-time downloads
58
Public
Parameters
324M
659 MB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors648 MB · 98%
From the Hugging Face model README
Fine-tune de Audio8/Audio8-ASR-0.1B especializado en español rioplatense / latinoamericano mediante SFT (supervised fine-tuning) completo.
El modelo base no incluía español entre sus 7 idiomas de entrenamiento (en, zh, fr, ja, yue, de, ko), por lo que su WER zero-shot en español rioplatense era ~65%. Tras el fine-tune baja a 12.39% WER / 5.25% CER.
| Modelo | WER | CER |
|---|---|---|
| Audio8/Audio8-ASR-0.1B (baseline, zero-shot) | 64.76% | 28.28% |
| Este modelo (es_google_v3) | 12.39% | 5.25% |
Idéntica al modelo base (324M parámetros end-to-end):
Se carga con trust_remote_code=True (custom code incluido en el repo).
~13.5k clips (~15h) de habla leída en español latinoamericano, combinando:
ylacombe/google-argentinian-spanish (female + male)ylacombe/google-chilean-spanish (female + male)ylacombe/google-colombian-spanish (female + male)Split 90/10 train/eval por dataset; el eval reportado es el held-out argentino.
Full fine-tune (todos los módulos descongelados, incluido el audio encoder), en 3 etapas con learning rate decreciente:
bf16, batch efectivo 32, cosine schedule, warmup 2-3%, weight decay 0.005, grad clip 1.0, AdamW. Audio cap 30s @ 16kHz. transformers 4.57.3.
import torch
from transformers import AutoModelForCausalLM, AutoProcessor
model_path = "surus-lat/audio8-asr-0.1b-es-rioplatense"
audio_path = "audio.wav"
device = "cuda" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.bfloat16 if device == "cuda" else torch.float32
processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
model_path, trust_remote_code=True,
torch_dtype=torch_dtype, attn_implementation="eager",
).to(device)
model.eval()
conversation = [{
"role": "user",
"content": [
{"type": "audio", "path": audio_path},
{"type": "text", "text": "Please transcribe this audio."},
],
}]
batch = processor.apply_chat_template(
conversation, return_tensors="pt", sampling_rate=16000,
audio_padding="longest", add_generation_prompt=True,
audio_max_length=30 * 16000,
text_kwargs={"padding": "longest", "truncation": True, "max_length": 1000},
)
batch = {k: (v.to(device) if hasattr(v, "to") else v) for k, v in dict(batch).items()}
with torch.inference_mode():
ids = model.generate(**batch, max_new_tokens=128, do_sample=False)
prompt_len = int(batch["input_ids"].shape[1])
print(processor.decode(ids[0, prompt_len:], skip_special_tokens=True).strip())
Importante: usar transformers==4.57.3. En transformers 5.x el lm_head
ligado (tied) no se carga correctamente y el modelo degrada a salida aleatoria.
Modelo base: Audio8-ASR-0.1B — Data-Efficient On-Policy Distillation for Automatic Speech Recognition (arXiv 2605.28139, AutoArk-AI).