如何使用NVIDIA Canary-1B-v2进行ASR、翻译和自动SRT字幕导出
阅读原文· marktechpost.com该教程通过Python使用NVIDIA Canary-1B-v2构建语音识别与翻译管线：先安装NeMo ASR工具包和音频依赖，在GPU上加载模型；将音频处理为16 kHz单声道格式后，执行英语ASR，将语音翻译为25种语言，生成词级和时间段时间戳，导出SRT字幕文件，并支持长文本转录、批处理与推理速度基准测试。
In this tutorial, we build a speech recognition and translation workflow using NVIDIA Canary-1B-v2. We begin by setting up the required audio, NeMo, NumPy, and SciPy dependencies, then load the Canary model on a GPU-enabled runtime for efficient inference. From there, we prepare audio into a clean 16 kHz mono format, perform English ASR, translate speech into multiple languages, generate word and segment timestamps, export translated subtitles as an SRT file, test long-form transcription, run batch processing, and benchmark inference speed. At the end, we have a complete multilingual ASR and speech translation pipeline that we can adapt for real audio files, subtitle generation, and large-scale transcription experiments.
Installing NeMo, Audio Libraries, NumPy, and SciPy Dependencies
import os, subprocess, sys
SENTINEL = "/content/.canary_setup_done"
if not os.path.exists(SENTINEL):
def sh(c):
print("$", c); subprocess.run(c, shell=True, check=False)
print(">>> PHASE 1: installing dependencies (one-time)...\n")
sh("apt-get -qq update")
sh("apt-get -qq install -y libsndfile1 ffmpeg > /dev/null")
sh('pip install -q "nemo_toolkit[asr]"')
sh("pip install -q librosa soundfile pydub")
sh('pip install -q --force-reinstall --no-cache-dir "numpy>=2.2,<2.4" "scipy>=1.15"')
open(SENTINEL, "w").write("done")
print("\n✅ Setup complete. Restarting the runtime now.")
print(" When it reconnects, RUN THIS CELL AGAIN to start the tutorial.")
os.kill(os.getpid(), 9)
We set up the environment for the NVIDIA Canary-1B-v2 tutorial. We install the required system packages, NeMo ASR toolkit, audio libraries, and compatible NumPy and SciPy versions. We then create a setup marker and restart the runtime so that the updated dependencies load cleanly before running the main tutorial.
Loading NVIDIA Canary-1B-v2 and Checking GPU Availability
import time, json, gc, math, urllib.request
import torch, numpy as np, soundfile as sf, librosa
print(">>> PHASE 2: running tutorial\n")
print("NumPy:", np.__version__, "| PyTorch:", torch.__version__)
print("CUDA available:", torch.cuda.is_available())
if torch.cuda.is_available():
print("GPU:", torch.cuda.get_device_name(0),
f"| VRAM: {torch.cuda.get_device_properties(0).total_memory/1e9:.1f} GB")
else:
print("⚠️ No GPU — will run on CPU (very slow). "
"Set Runtime > Change runtime type > GPU.")
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
LANGS = {
"bg":"Bulgarian","hr":"Croatian","cs":"Czech","da":"Danish","nl":"Dutch",
"en":"English","et":"Estonian","fi":"Finnish","fr":"French","de":"German",
"el":"Greek","hu":"Hungarian","it":"Italian","lv":"Latvian","lt":"Lithuanian",
"mt":"Maltese","pl":"Polish","pt":"Portuguese","ro":"Romanian","sk":"Slovak",
"sl":"Slovenian","es":"Spanish","sv":"Swedish","ru":"Russian","uk":"Ukrainian",
}
print(f"\nSupported languages ({len(LANGS)}):", ", ".join(LANGS.keys()))
from nemo.collections.asr.models import ASRModel
print("\nLoading nvidia/canary-1b-v2 ...")
t0 = time.time()
asr_model = ASRModel.from_pretrained(model_name="nvidia/canary-1b-v2").to(DEVICE).eval()
print(f"Model loaded in {time.time()-t0:.1f}s")
We import the main libraries and check whether CUDA is available for GPU acceleration. We define the supported language dictionary to enable Canary to handle multilingual ASR and translation tasks. We then load the NVIDIA Canary-1B-v2 model from NeMo and move it to the available device for inference.
Preparing 16 kHz Audio and Running English ASR with Translation
TARGET_SR = 16000
def prepare_audio(path_or_url, out_path=None):
if str(path_or_url).startswith(("http://", "https://")):
local = "/content/_dl_" + os.path.basename(path_or_url.split("?")[0])
urllib.request.urlretrieve(path_or_url, local)
path_or_url = local
audio, _ = librosa.load(path_or_url, sr=TARGET_SR, mono=True)
if out_path is None:
base = os.path.splitext(os.path.basename(path_or_url))[0]
out_path = f"/content/{base}_16k_mono.wav"
sf.write(out_path, audio, TARGET_SR, subtype="PCM_16")
dur = len(audio) / TARGET_SR
print(f"Prepared: {out_path} ({dur:.1f}s, 16kHz, mono)")
return out_path, dur
SAMPLE_URL = "https://dldata-public.s3.us-east-2.amazonaws.com/2086-149220-0033.wav"
sample_wav, sample_dur = prepare_audio(SAMPLE_URL)
def transcribe(files, source_lang="en", target_lang="en", timestamps=False, batch_size=1):
if isinstance(files, str):
files = [files]
return asr_model.transcribe(files, source_lang=source_lang, target_lang=target_lang,
timestamps=timestamps, batch_size=batch_size)
print("\n=== 1) BASIC ASR (English) ===")
res = transcribe(sample_wav, source_lang="en", target_lang="en")
print("Transcript:", res[0].text)
print("\n=== 2) TRANSLATION (EN audio -> X) ===")
for tgt in ["fr", "de", "es", "it"]:
out = transcribe(sample_wav, source_lang="en", target_lang=tgt)
print(f" EN -> {LANGS[tgt]:<10} ({tgt}): {out[0].text}")
We create a reusable audio preparation function that downloads audio when needed and converts it into 16 kHz mono WAV format. We load the sample audio file and define a helper function for transcription and translation. We then run basic English ASR and translate the same English speech into French, German, Spanish, and Italian.
Generating Word and Segment Timestamps and Exporting SRT Subtitles
print("\n=== 3) TIMESTAMPS (ASR) ===")
ts_out = transcribe(sample_wav, source_lang="en", target_lang="en", timestamps=True)
word_ts = ts_out[0].timestamp.get("word", [])
seg_ts = ts_out[0].timestamp.get("segment", [])
print("Segments:")
for s in seg_ts:
print(f" [{s['start']:6.2f}s - {s['end']:6.2f}s] {s['segment']}")
print("First 10 words:")
for w in word_ts[:10]:
print(f" [{w['start']:6.2f}s - {w['end']:6.2f}s] {w['word']}")
def _srt_time(t):
h=int(t//3600); m=int((t%3600)//60); s=int(t%60); ms=int(round((t-int(t))*1000))
return f"{h:02d}:{m:02d}:{s:02d},{ms:03d}"
def segments_to_srt(segments, out_path="/content/output.srt"):
lines=[]
for i, seg in enumerate(segments, 1):
lines += [str(i), f"{_srt_time(seg['start'])} --> {_srt_time(seg['end'])}",
seg["segment"].strip(), ""]
open(out_path, "w", encoding="utf-8").write("\n".join(lines))
print(f"Saved SRT: {out_path}")
return out_path
print("\n=== 4) SRT EXPORT (translated French subtitles) ===")
fr_ts = transcribe(sample_wav, source_lang="en", target_lang="fr", timestamps=True)
segments_to_srt(fr_ts[0].timestamp["segment"], "/content/subtitles_fr.srt")
print(open("/content/subtitles_fr.srt").read())
We enable timestamped transcription to extract both segment-level and word-level timing information. We print the transcript segments and the first few word timestamps to inspect how the model aligns text with audio. We also convert translated French segments into an SRT subtitle file and display the generated subtitles.
Running Long-Form Transcription, Batch Processing, and Speed Benchmark
print("\n=== 5) LONG-FORM (sample tiled x6) ===")
long_audio, _ = librosa.load(sample_wav, sr=TARGET_SR, mono=True)
long_audio = np.tile(long_audio, 6)
sf.write("/content/long.wav", long_audio, TARGET_SR, subtype="PCM_16")
print(f"Long clip duration: {len(long_audio)/TARGET_SR:.1f}s")
long_out = transcribe("/content/long.wav", source_lang="en", target_lang="en", batch_size=1)
print("Long transcript (first 300 chars):", long_out[0].text[:300], "...")
print("\n=== 6) BATCH ===")
for name in ["clip_a", "clip_b"]:
sf.write(f"/content/{name}.wav",
librosa.load(sample_wav, sr=TARGET_SR, mono=True)[0], TARGET_SR, subtype="PCM_16")
batch = transcribe(["/content/clip_a.wav", "/content/clip_b.wav"],
source_lang="en", target_lang="en", batch_size=2)
for i, b in enumerate(batch):
print(f" file {i}: {b.text}")
print("\n=== 7) BENCHMARK ===")
t0 = time.time(); _ = transcribe(sample_wav, source_lang="en", target_lang="en")
elapsed = time.time()-t0
print(f"Audio: {sample_dur:.2f}s | Compute: {elapsed:.2f}s | RTFx ≈ {sample_dur/elapsed:.1f}x")
print("\n✅ Done. Change source_lang/target_lang from the LANGS dict to try other languages.")
We test long-form transcription by repeating the sample audio several times and passing the longer clip through the model. We also create two duplicate audio clips to demonstrate batch transcription with a batch size of two. Also, we benchmark the model by comparing audio duration with compute time and report the real-time factor speed.
Conclusion
In conclusion, we completed a practical end-to-end workflow for using NVIDIA Canary-1B-v2 as a multilingual ASR and speech translation system. We processed raw audio, generated accurate transcripts, translated speech into different target languages, extracted timestamps, created subtitle files, handled longer audio clips, and compared runtime performance through a simple benchmark. We now have a reusable Colab-ready pipeline that we can extend further with custom uploads, more languages, larger batches, and production-style audio processing.