Spaces:
Running
Running
Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
import torchaudio
|
| 3 |
+
import gradio as gr
|
| 4 |
+
from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq
|
| 5 |
+
|
| 6 |
+
# device setup
|
| 7 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 8 |
+
|
| 9 |
+
# load model + processor
|
| 10 |
+
model_name = "ibm-granite/granite-speech-3.3-8b"
|
| 11 |
+
processor = AutoProcessor.from_pretrained(model_name)
|
| 12 |
+
tokenizer = processor.tokenizer
|
| 13 |
+
model = AutoModelForSpeechSeq2Seq.from_pretrained(
|
| 14 |
+
model_name, device_map=device, torch_dtype=torch.bfloat16
|
| 15 |
+
)
|
| 16 |
+
|
| 17 |
+
today_str = date.today().strftime("%B %d, %Y")
|
| 18 |
+
|
| 19 |
+
system_prompt = (
|
| 20 |
+
"Knowledge Cutoff Date: April 2024.\n"
|
| 21 |
+
f"Today's Date: {today_str}.\n"
|
| 22 |
+
"You are Granite, developed by IBM. You are a helpful AI assistant."
|
| 23 |
+
)
|
| 24 |
+
|
| 25 |
+
def transcribe(audio_file):
|
| 26 |
+
# load wav file
|
| 27 |
+
wav, sr = torchaudio.load(audio_file, normalize=True)
|
| 28 |
+
if wav.shape[0] != 1 or sr != 16000:
|
| 29 |
+
# resample + convert to mono if needed
|
| 30 |
+
wav = torch.mean(wav, dim=0, keepdim=True) # mono
|
| 31 |
+
wav = torchaudio.functional.resample(wav, sr, 16000)
|
| 32 |
+
sr = 16000
|
| 33 |
+
|
| 34 |
+
# user prompt
|
| 35 |
+
user_prompt = "<|audio|>can you transcribe the speech into a written format?"
|
| 36 |
+
chat = [
|
| 37 |
+
dict(role="system", content=system_prompt),
|
| 38 |
+
dict(role="user", content=user_prompt),
|
| 39 |
+
]
|
| 40 |
+
prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
|
| 41 |
+
|
| 42 |
+
# run model
|
| 43 |
+
model_inputs = processor(prompt, wav, sampling_rate=sr, device=device, return_tensors="pt").to(device)
|
| 44 |
+
model_outputs = model.generate(
|
| 45 |
+
**model_inputs,
|
| 46 |
+
max_new_tokens=200,
|
| 47 |
+
do_sample=False,
|
| 48 |
+
num_beams=1
|
| 49 |
+
)
|
| 50 |
+
|
| 51 |
+
# strip prompt tokens
|
| 52 |
+
num_input_tokens = model_inputs["input_ids"].shape[-1]
|
| 53 |
+
new_tokens = torch.unsqueeze(model_outputs[0, num_input_tokens:], dim=0)
|
| 54 |
+
output_text = tokenizer.batch_decode(
|
| 55 |
+
new_tokens, add_special_tokens=False, skip_special_tokens=True
|
| 56 |
+
)
|
| 57 |
+
|
| 58 |
+
return output_text[0].strip()
|
| 59 |
+
|
| 60 |
+
# Gradio UI
|
| 61 |
+
with gr.Blocks() as demo:
|
| 62 |
+
gr.Markdown("## Granite 3.3 Speech-to-Text Demo")
|
| 63 |
+
|
| 64 |
+
with gr.Row():
|
| 65 |
+
audio_input = gr.Audio(type="filepath", label="Upload Audio (16kHz mono preferred)")
|
| 66 |
+
output_text = gr.Textbox(label="Transcription", lines=5)
|
| 67 |
+
|
| 68 |
+
transcribe_btn = gr.Button("Transcribe")
|
| 69 |
+
transcribe_btn.click(fn=transcribe, inputs=audio_input, outputs=output_text)
|
| 70 |
+
|
| 71 |
+
demo.launch()
|