Spaces:
Runtime error
Runtime error
Commit
·
5bc92c5
1
Parent(s):
e46b406
Fix pipeline error by using model.generate() directlyclear
Browse files
app.py
CHANGED
|
@@ -1,30 +1,85 @@
|
|
| 1 |
import gradio as gr
|
| 2 |
-
from transformers import pipeline
|
| 3 |
import torch
|
| 4 |
-
from nemo.collections.speechlm2 import SALM
|
| 5 |
import spaces
|
|
|
|
|
|
|
| 6 |
|
| 7 |
-
if torch.cuda.is_available()
|
| 8 |
-
|
| 9 |
-
else:
|
| 10 |
-
device = torch.device("cpu")
|
| 11 |
-
|
| 12 |
-
SAMPLE_RATE = 16000 # Hz - NVIDIA model sampling rate
|
| 13 |
-
MAX_AUDIO_MINUTES = 120 # wont try to transcribe if longer than this
|
| 14 |
-
CHUNK_SECONDS = 40.0 # max audio length seen by the model
|
| 15 |
-
BATCH_SIZE = 192 # for parallel transcription of audio longer than CHUNK_SECONDS
|
| 16 |
|
| 17 |
-
# Initialize the ASR model which is based on the "nvidia/canary-qwen-2.5b" architecture and uses NVIDIA's NeMo framework
|
| 18 |
model = SALM.from_pretrained("nvidia/canary-qwen-2.5b").bfloat16().eval().to(device)
|
| 19 |
-
transcriber = pipeline("automatic-speech-recognition", model = model)
|
| 20 |
|
| 21 |
-
# Transcribe audio file using NeMo's transcribe class and use spaces for GPU acceleration
|
| 22 |
@spaces.GPU
|
| 23 |
-
def transcribe_audio(
|
| 24 |
-
|
| 25 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
|
| 27 |
-
demo
|
| 28 |
-
|
| 29 |
-
inputs=gr.Audio(source="upload", type="filepath"),
|
| 30 |
-
outputs=gr.Textbox())
|
|
|
|
| 1 |
import gradio as gr
|
|
|
|
| 2 |
import torch
|
|
|
|
| 3 |
import spaces
|
| 4 |
+
from lhotse import Recording
|
| 5 |
+
from nemo.collections.speechlm2 import SALM
|
| 6 |
|
| 7 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 8 |
+
SAMPLE_RATE = 16000
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
|
|
|
|
| 10 |
model = SALM.from_pretrained("nvidia/canary-qwen-2.5b").bfloat16().eval().to(device)
|
|
|
|
| 11 |
|
|
|
|
| 12 |
@spaces.GPU
|
| 13 |
+
def transcribe_audio(audio_filepath):
|
| 14 |
+
if audio_filepath is None:
|
| 15 |
+
return "Please upload an audio file", ""
|
| 16 |
+
|
| 17 |
+
rec = Recording.from_file(audio_filepath, recording_id="temp")
|
| 18 |
+
cut = rec.resample(SAMPLE_RATE).to_cut()
|
| 19 |
+
if cut.num_channels > 1:
|
| 20 |
+
cut = cut.to_mono(mono_downmix=True)
|
| 21 |
+
|
| 22 |
+
audio, audio_lens = cut.load_audio()
|
| 23 |
+
|
| 24 |
+
with torch.inference_mode():
|
| 25 |
+
output_ids = model.generate(
|
| 26 |
+
prompts=[[{"role": "user", "content": f"Transcribe the following: {model.audio_locator_tag}"}]],
|
| 27 |
+
audios=torch.as_tensor(audio).unsqueeze(0).to(device),
|
| 28 |
+
audio_lens=torch.as_tensor([audio_lens]).to(device),
|
| 29 |
+
max_new_tokens=256,
|
| 30 |
+
)
|
| 31 |
+
|
| 32 |
+
transcript = model.tokenizer.ids_to_text(output_ids[0].cpu())
|
| 33 |
+
return transcript, transcript
|
| 34 |
+
|
| 35 |
+
@spaces.GPU
|
| 36 |
+
def answer_question(transcript, question):
|
| 37 |
+
if not transcript:
|
| 38 |
+
return "Please transcribe audio first"
|
| 39 |
+
|
| 40 |
+
with torch.inference_mode(), model.llm.disable_adapter():
|
| 41 |
+
output_ids = model.generate(
|
| 42 |
+
prompts=[[{"role": "user", "content": f"{question}\n\n{transcript}"}]],
|
| 43 |
+
max_new_tokens=512,
|
| 44 |
+
)
|
| 45 |
+
|
| 46 |
+
answer = model.tokenizer.ids_to_text(output_ids[0].cpu())
|
| 47 |
+
answer = answer.split("<|im_start|>assistant")[-1]
|
| 48 |
+
return answer.strip()
|
| 49 |
+
|
| 50 |
+
with gr.Blocks(title="Canary-Qwen Transcriber & Q&A") as demo:
|
| 51 |
+
gr.Markdown("# Canary-Qwen Transcriber with Q&A")
|
| 52 |
+
gr.Markdown("Upload audio to transcribe, then ask questions about it!")
|
| 53 |
+
|
| 54 |
+
with gr.Row():
|
| 55 |
+
with gr.Column():
|
| 56 |
+
audio_input = gr.Audio(sources=["microphone", "upload"], type="filepath", label="Audio Input")
|
| 57 |
+
transcribe_btn = gr.Button("Transcribe", variant="primary")
|
| 58 |
+
|
| 59 |
+
with gr.Column():
|
| 60 |
+
transcript_output = gr.Textbox(label="Transcript", lines=8)
|
| 61 |
+
|
| 62 |
+
transcript_state = gr.State()
|
| 63 |
+
|
| 64 |
+
with gr.Row():
|
| 65 |
+
with gr.Column():
|
| 66 |
+
question_input = gr.Textbox(label="Ask a question about the transcript", placeholder="What is the main topic?")
|
| 67 |
+
ask_btn = gr.Button("Ask", variant="primary")
|
| 68 |
+
|
| 69 |
+
with gr.Column():
|
| 70 |
+
answer_output = gr.Textbox(label="Answer", lines=4)
|
| 71 |
+
|
| 72 |
+
transcribe_btn.click(
|
| 73 |
+
fn=transcribe_audio,
|
| 74 |
+
inputs=[audio_input],
|
| 75 |
+
outputs=[transcript_output, transcript_state]
|
| 76 |
+
)
|
| 77 |
+
|
| 78 |
+
ask_btn.click(
|
| 79 |
+
fn=answer_question,
|
| 80 |
+
inputs=[transcript_state, question_input],
|
| 81 |
+
outputs=[answer_output]
|
| 82 |
+
)
|
| 83 |
|
| 84 |
+
demo.queue()
|
| 85 |
+
demo.launch()
|
|
|
|
|
|