import os
os.environ["TOKENIZERS_PARALLELISM"] = "false"
import gradio as gr
import torch
import spaces
import torchaudio
import uuid
import time
from datetime import timedelta
from lhotse import Recording
from lhotse.dataset import DynamicCutSampler
from nemo.collections.speechlm2 import SALM
from pathlib import Path
# Set synthwave theme
theme = gr.themes.Ocean(
primary_hue="indigo",
secondary_hue="fuchsia",
neutral_hue="slate",
).set(
button_large_radius='*radius_sm'
)
# Set device to use cuda if available and sample rate to 16000 for Nvidia NeMo
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
SAMPLE_RATE = 16000
MAX_AUDIO_MINUTES = 120
CHUNK_SECONDS = 40.0
BATCH_SIZE = 192
# Load the model from Hugging Face Hub using Nvidia SALM
model = SALM.from_pretrained("nvidia/canary-qwen-2.5b").bfloat16().eval().to(device)
def as_batches(audio_filepath, utt_id):
rec = Recording.from_file(audio_filepath, recording_id=utt_id)
if rec.duration / 60.0 > MAX_AUDIO_MINUTES:
raise gr.Error(f"Audio file is too long. Maximum duration is {MAX_AUDIO_MINUTES} minutes.")
cut = rec.resample(SAMPLE_RATE).to_cut()
if cut.num_channels > 1:
cut = cut.to_mono(mono_downmix=True)
return DynamicCutSampler(cut.cut_into_windows(CHUNK_SECONDS), max_cuts=BATCH_SIZE)
# Define the audio transcription function and use ZeroGPU
@spaces.GPU
def transcribe_audio(audio_filepath):
if audio_filepath is None:
return "Please upload an audio file", "", [], ""
start_time = time.time()
utt_id = uuid.uuid4()
pred_text = []
for batch in as_batches(audio_filepath, str(utt_id)):
audio, audio_lens = batch.load_audio(collate=True)
with torch.inference_mode():
output_ids = model.generate(
prompts=[[{"role": "user", "content": f"Transcribe the following using accurate punctuation and capitalization: {model.audio_locator_tag}"}]] * len(batch),
audios=torch.as_tensor(audio).to(device, non_blocking=True),
audio_lens=torch.as_tensor(audio_lens).to(device, non_blocking=True),
max_new_tokens=256,
)
texts = [model.tokenizer.ids_to_text(oids) for oids in output_ids.cpu()]
for t in texts:
pred_text.append(t)
transcript = ' '.join(pred_text)
end_time = time.time()
# Calculate statistics
transcription_time = round(end_time - start_time, 2)
word_count = len(transcript.split())
words_per_second = round(word_count / transcription_time, 2) if transcription_time > 0 else 0
# Get filename
filename = Path(audio_filepath).name
# Create label with stats
label_text = f"File: {filename} | Words: {word_count} | Time: {transcription_time}s | WPS: {words_per_second}"
return transcript, transcript, gr.update(label=label_text)
@spaces.GPU
def transcript_qa(transcript, question):
if not transcript:
return "Please transcribe audio first before asking questions.", ""
if not question or question.strip() == "":
return "", ""
with torch.inference_mode(), model.llm.disable_adapter():
output_ids = model.generate(
prompts=[[{"role": "user", "content": f"{question}\n\nTranscript: {transcript}"}]],
max_new_tokens=1024,
)
ans = model.tokenizer.ids_to_text(output_ids[0].cpu())
ans = ans.split("<|im_start|>assistant")[-1]
thinking = ""
if "" in ans:
if "" in ans:
parts = ans.split("")
# Get text before tag if any
before_think = parts[0] if len(parts) > 1 else ""
# Get content between and
think_content = parts[1] if len(parts) > 1 else parts[0]
thinking, after_think = think_content.split("")
thinking = thinking.strip()
# Combine text before and after thinking
ans = before_think + after_think
ans = ans.strip()
if not ans:
ans = "I couldn't generate a response. Please try rephrasing your question."
return ans, thinking
def disable_transcribe():
return gr.update(interactive=False)
def enable_transcribe():
return gr.update(interactive=True)
# Load external CSS and HTML
def load_template(filename):
template_path = Path(__file__).parent / "templates" / filename
return template_path.read_text() if template_path.exists() else ""
# Build the Gradio interface
with gr.Blocks(theme=theme) as demo:
# Simple banner image - responsive and clean
gr.HTML("""
""")
gr.Markdown("## Upload an Audio File, Choose an Example File, or Record Yourself Then Ask Questions About the Transcript.")
gr.Markdown('''NVIDIA NeMo Canary-Qwen-2.5B is an English speech recognition model that achieves state-of-the art
performance on multiple English speech benchmarks. With 2.5 billion parameters and running at 418 RTFx,
Canary-Qwen-2.5B supports automatic speech-to-text recognition (ASR) in English with punctuation and capitalization
(PnC). The model works in two modes: as a transcription tool (ASR mode) and as an LLM (LLM mode). In ASR mode, the
model is only capable of transcribing the speech into text, but does not retain any LLM-specific skills such as reasoning.
In LLM mode, the model retains all of the original LLM capabilities, which can be used to post-process the transcript, e.g.
summarize it or answer questions about it. In LLM mode, the model does not "understand" the raw audio anymore - only
its transcript. This model is ready for commercial use. All example audio was generated using Microsoft VibeVoice, found in my other space
- [Conference Generator VibeVoice](https://huggingface.co/spaces/ACloudCenter/Conference-Generator-VibeVoice)''')
with gr.Tabs():
with gr.Tab("Transcribe"):
# State variables
transcript_state = gr.State("")
# Example questions
example_questions = [
["Can you summarize this meeting?"],
["Please provide bullet points of the key items."],
["What is the TL;DR of this meeting?"],
["What was the main take-away?"],
["What was the main topic?"],
]
# Define file paths as variables
ai_ted = "public/audio_files/ai_tedtalk.wav"
financial = "public/audio_files/financial_meeting.wav"
military = "public/audio_files/military_meeting.wav"
oil = "public/audio_files/oil_meeting.wav"
political = "public/audio_files/political_speech.wav"
telehealth = "public/audio_files/telehealth_meeting.wav"
game_dev = "public/audio_files/game_create_meeting.wav"
product = "public/audio_files/product_meeting.wav"
# Audio Input and Transcript
with gr.Row():
with gr.Column(scale=1):
gr.Markdown("### Audio Input")
audio_input = gr.Audio(
sources=["microphone", "upload"],
type="filepath",
label="Record/Upload Audio (MP3, WAV, M4A, etc.)",
show_download_button=True
)
gr.Examples(
examples=[
[ai_ted],
[financial],
[military],
[oil],
[political],
[telehealth],
[game_dev],
[product]
],
inputs=audio_input,
label="Example Audio Files",
example_labels=["AI TED Talk", "Financial Meeting", "Military Meeting", "Oil & Gas Meeting",
"Political Speech", "Telehealth Meeting", "Game Dev Meeting", "Product Meeting"]
)
transcribe_btn = gr.Button("Transcribe Audio", variant="primary", size="lg")
clear_audio_btn = gr.Button("Clear Audio")
with gr.Column(scale=1):
gr.Markdown("### Transcript")
transcript_output = gr.Textbox(
label="Waiting for transcription...",
lines=12,
placeholder="Transcript will appear here after clicking 'Transcribe Audio'...",
max_lines=12,
autoscroll=True
)
clear_transcript_btn = gr.Button("Clear Transcript")
# Spacing
gr.Markdown("---")
with gr.Row():
with gr.Column(scale=1):
gr.Markdown("### Interactive Q&A")
gr.Markdown("#### About Context-Aware Q&A")
gr.Markdown("""The model retains the full transcript context, allowing you to ask follow-up questions
naturally without re-stating information. It understands references like 'they', 'it', or 'that topic'.""")
gr.Markdown("#### Example Questions")
# Examples will be added after msg is defined
example_container = gr.Column()
with gr.Column(scale=3):
# Add thinking display above chat
with gr.Accordion("🧠Model Thinking", open=False):
thinking_box = gr.Textbox(
label="",
placeholder="The model's reasoning will appear here when available...",
lines=6,
max_lines=10,
interactive=False
)
chatbot = gr.Chatbot(
label="Response",
type="messages",
height=400,
show_copy_button=True,
autoscroll=True
)
with gr.Row():
msg = gr.Textbox(
placeholder="Ask a question about the transcript...",
label="Your Question",
lines=1
)
submit_chat_btn = gr.Button("Send", variant="primary", scale=1)
clear_chat_btn = gr.Button("Clear Chat", size="sm")
# Event handlers
def submit_question(question, transcript):
if not question or question.strip() == "":
yield "", [], ""
answer, thinking = transcript_qa(transcript, question)
# Just show the current Q&A, no history
messages = [
{"role": "user", "content": question},
{"role": "assistant", "content": answer}
]
yield "", messages, thinking
# Add examples inside the left column container
with example_container:
gr.Examples(
examples=example_questions,
inputs=msg,
outputs=[msg, chatbot, thinking_box],
fn=lambda q: submit_question(q, transcript_state.value),
cache_examples=False,
label=""
)
transcribe_btn.click(
fn=disable_transcribe,
outputs=[transcribe_btn]
).then(
fn=lambda: ([], ""),
outputs=[chatbot, thinking_box]
).then(
fn=transcribe_audio,
inputs=[audio_input],
outputs=[transcript_output, transcript_state, transcript_output] # Third output updates the label
).then(
fn=enable_transcribe,
outputs=[transcribe_btn]
)
clear_audio_btn.click(
fn=lambda: None,
outputs=[audio_input]
)
clear_transcript_btn.click(
fn=lambda: ("", "", gr.update(label="Waiting for transcription...")),
outputs=[transcript_output, transcript_state, transcript_output]
)
msg.submit(
fn=submit_question,
inputs=[msg, transcript_state],
outputs=[msg, chatbot, thinking_box]
)
submit_chat_btn.click(
fn=submit_question,
inputs=[msg, transcript_state],
outputs=[msg, chatbot, thinking_box]
)
clear_chat_btn.click(
fn=lambda: ([], ""),
outputs=[chatbot, thinking_box]
)
with gr.Tab("Architecture"):
gr.Markdown("### Model Performance")
with gr.Row():
with gr.Column(scale=1):
gr.Markdown("""
#### Industry-Leading Performance
Canary ranks at the top of the HuggingFace Open ASR Leaderboard with an average word error rate (WER) of **6.67%**. It outperforms all other open-source models by a wide margin.
#### Training Data
Canary is trained on a combination of public and in-house data:
- **85K hours** of transcribed speech for speech recognition
- NVIDIA NeMo text translation models used to generate translations of the original transcripts in all supported languages
Despite using an order of magnitude less data, Canary outperforms the similarly sized Whisper-large-v3 and SeamlessM4T-Medium-v1 models on both transcription and translation tasks.
""")
gr.Markdown("### Benchmark Results")
gr.Markdown("""
#### Word Error Rate (WER) on MCV 16.1 Test Sets
On the MCV 16.1 test sets for English, Spanish, French, and German, Canary achieved a WER of **5.77** (lower is better).
""")
with gr.Column(scale=3):
gr.HTML("""
NVIDIA ASR
""")
with gr.Row():
with gr.Column(scale=1):
gr.Markdown("""
| Model | Average WER |
|-------|-------------|
| **Canary** | **5.77** |
| SeamlessM4T-v2 | 6.41 |
| Whisper-large-v3 | 8.05 |
| SeamlessM4T-v1 | 9.48 |
""")
with gr.Column(scale=3):
gr.Markdown("""
#### Translation BLEU Scores
**From English** (ES, FR, DE on FLEURS & MExpresso):
- Canary: **30.57** BLEU
**To English** (ES, FR, DE on FLEURS & CoVoST):
- Canary: **34.25** BLEU
*(Higher BLEU scores indicate better translation quality)*
""")
gr.Markdown("---")
with gr.Row():
with gr.Column():
gr.Markdown("""
### Canary Architecture Details
Canary is an encoder-decoder model built on NVIDIA innovations:
- **Encoder**: Fast-Conformer - an efficient Conformer architecture optimized for ~3x savings on compute and ~4x savings on memory
- **Processing**: Audio is processed as log-mel spectrogram features
- **Decoder**: Transformer decoder generates output text tokens auto-regressively
- **Control**: Special tokens control whether Canary performs transcription or translation
- **Tokenizer**: Concatenated tokenizer offers explicit control of output token space
#### Licensing
- **Model weights**: CC BY-NC 4.0 license (research-friendly, non-commercial)
- **Training code**: Apache 2.0 license (available from NeMo)
For more information about accessing Canary locally and building on top of it, see the [NVIDIA/NeMo GitHub repository](https://github.com/NVIDIA/NeMo).
""")
with gr.Column(scale=3):
gr.HTML("""
