whisper.youtube: sliding-window streaming + live mic mode for Parakeet
- parakeet_transcribe.py: rewrite --stream as sliding window with overlap dedup (case-/punctuation-insensitive); add --window and --shift params - parakeet_mic.py: live mic capture via pw-record/arecord, progressive transcription with same sliding-window algorithm - parakeet: export NUMBA_CACHE_DIR to work around librosa caching error on the mounted filesystem Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
nnnet
parent
7a4ff7ec36
commit
3ea11b4e5b
|
|
@ -5,4 +5,8 @@
|
|||
# ./parakeet <youtube-url>
|
||||
set -euo pipefail
|
||||
DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
|
||||
# numba cannot write its cache on the mounted filesystem under this dir
|
||||
# (no inode-stable locator); redirect cache to /tmp.
|
||||
export NUMBA_CACHE_DIR="${NUMBA_CACHE_DIR:-/tmp/numba-cache-$USER}"
|
||||
mkdir -p "$NUMBA_CACHE_DIR"
|
||||
exec "$DIR/.venv-parakeet/bin/python" "$DIR/parakeet_transcribe.py" "$@"
|
||||
|
|
|
|||
|
|
@ -0,0 +1,152 @@
|
|||
#!/usr/bin/env python3
|
||||
"""Live microphone streaming with Parakeet-tdt-0.6b-v3.
|
||||
|
||||
Captures audio from the mic (pw-record or arecord), runs a sliding-window
|
||||
transcription, and prints text progressively as new audio arrives.
|
||||
"""
|
||||
|
||||
import argparse
|
||||
import os
|
||||
import shutil
|
||||
import signal
|
||||
import subprocess
|
||||
import sys
|
||||
import tempfile
|
||||
import time
|
||||
|
||||
import numpy as np
|
||||
|
||||
MODEL_NAME = "nvidia/parakeet-tdt-0.6b-v3"
|
||||
SR = 16000
|
||||
DEFAULT_WINDOW_S = 8.0
|
||||
DEFAULT_SHIFT_S = 2.5
|
||||
|
||||
|
||||
def _norm(w: str) -> str:
|
||||
return "".join(ch for ch in w.lower() if ch.isalnum())
|
||||
|
||||
|
||||
def _dedup_overlap(prev_text: str, curr_text: str) -> str:
|
||||
prev_words = prev_text.split()
|
||||
curr_words = curr_text.split()
|
||||
prev_norm = [_norm(w) for w in prev_words]
|
||||
curr_norm = [_norm(w) for w in curr_words]
|
||||
max_n = min(len(prev_norm), len(curr_norm))
|
||||
for n in range(max_n, 0, -1):
|
||||
if prev_norm[-n:] == curr_norm[:n]:
|
||||
return " ".join(curr_words[n:])
|
||||
return curr_text
|
||||
|
||||
|
||||
def _find_recorder():
|
||||
for p in ("pw-record", "/usr/bin/pw-record", "/usr/local/bin/pw-record"):
|
||||
if shutil.which(p) if not os.path.isabs(p) else os.path.isfile(p):
|
||||
return p
|
||||
return shutil.which("arecord")
|
||||
|
||||
|
||||
def _build_record_cmd():
|
||||
rec = _find_recorder()
|
||||
if not rec:
|
||||
raise RuntimeError("Neither pw-record nor arecord found")
|
||||
if "pw-record" in rec:
|
||||
return [rec, "--format", "s16", "--rate", str(SR), "--channels", "1", "-"]
|
||||
return [rec, "-f", "S16_LE", "-r", str(SR), "-c", "1", "-t", "raw"]
|
||||
|
||||
|
||||
def load_model():
|
||||
import nemo.collections.asr as nemo_asr
|
||||
from nemo.utils import logging as nemo_logging
|
||||
import torch
|
||||
import logging
|
||||
|
||||
nemo_logging.set_verbosity(nemo_logging.ERROR)
|
||||
logging.getLogger("nemo_logger").setLevel(logging.ERROR)
|
||||
print(f"Loading {MODEL_NAME} ...", flush=True)
|
||||
asr = nemo_asr.models.ASRModel.from_pretrained(MODEL_NAME)
|
||||
asr = asr.cuda() if torch.cuda.is_available() else asr
|
||||
nemo_logging.set_verbosity(nemo_logging.ERROR)
|
||||
return asr
|
||||
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser(description=__doc__.splitlines()[0])
|
||||
parser.add_argument("--window", type=float, default=DEFAULT_WINDOW_S,
|
||||
help=f"sliding window length in seconds (default: {DEFAULT_WINDOW_S})")
|
||||
parser.add_argument("--shift", type=float, default=DEFAULT_SHIFT_S,
|
||||
help=f"window shift / step in seconds (default: {DEFAULT_SHIFT_S})")
|
||||
args = parser.parse_args()
|
||||
|
||||
if args.shift <= 0 or args.window <= 0:
|
||||
sys.exit("window and shift must be positive")
|
||||
if args.shift > args.window:
|
||||
sys.exit("shift must be <= window")
|
||||
|
||||
import soundfile as sf
|
||||
|
||||
asr = load_model()
|
||||
cmd = _build_record_cmd()
|
||||
|
||||
win_samples = int(args.window * SR)
|
||||
shift_samples = int(args.shift * SR)
|
||||
shift_bytes = shift_samples * 2 # int16 mono
|
||||
|
||||
tmp = tempfile.NamedTemporaryFile(suffix=".wav", prefix="parakeet-mic-", delete=False)
|
||||
tmp_wav = tmp.name
|
||||
tmp.close()
|
||||
|
||||
print(f"Recorder: {cmd[0]}", flush=True)
|
||||
print(f"Window: {args.window:.1f}s, shift: {args.shift:.1f}s "
|
||||
f"(overlap: {args.window - args.shift:.1f}s)", flush=True)
|
||||
print("--- speak into mic; Ctrl-C to stop ---\n", flush=True)
|
||||
|
||||
proc = subprocess.Popen(cmd, stdout=subprocess.PIPE, stderr=subprocess.DEVNULL)
|
||||
|
||||
def _shutdown(*_):
|
||||
if proc.poll() is None:
|
||||
proc.send_signal(signal.SIGINT)
|
||||
try:
|
||||
proc.wait(timeout=2)
|
||||
except subprocess.TimeoutExpired:
|
||||
proc.kill()
|
||||
if os.path.exists(tmp_wav):
|
||||
os.unlink(tmp_wav)
|
||||
print("\n--- stopped ---", flush=True)
|
||||
sys.exit(0)
|
||||
|
||||
signal.signal(signal.SIGINT, _shutdown)
|
||||
signal.signal(signal.SIGTERM, _shutdown)
|
||||
|
||||
buffer = np.zeros(0, dtype=np.int16)
|
||||
prev_text = ""
|
||||
|
||||
try:
|
||||
while True:
|
||||
new_bytes = proc.stdout.read(shift_bytes)
|
||||
if not new_bytes:
|
||||
break
|
||||
new_samples = np.frombuffer(new_bytes, dtype=np.int16)
|
||||
buffer = np.concatenate([buffer, new_samples])
|
||||
if len(buffer) > win_samples:
|
||||
buffer = buffer[-win_samples:]
|
||||
if len(buffer) < shift_samples:
|
||||
continue
|
||||
|
||||
sf.write(tmp_wav, buffer, SR)
|
||||
t0 = time.monotonic()
|
||||
out = asr.transcribe([tmp_wav], timestamps=False, verbose=False)
|
||||
elapsed = time.monotonic() - t0
|
||||
current = out[0].text.strip()
|
||||
if not current:
|
||||
continue
|
||||
|
||||
new_part = _dedup_overlap(prev_text, current) if prev_text else current
|
||||
if new_part.strip():
|
||||
print(new_part, end=" ", flush=True)
|
||||
prev_text = current
|
||||
finally:
|
||||
_shutdown()
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
|
|
@ -1,14 +1,28 @@
|
|||
#!/usr/bin/env python3
|
||||
"""Transcribe audio with NVIDIA Parakeet-tdt-0.6b-v3 (multilingual)."""
|
||||
"""Transcribe audio with NVIDIA Parakeet-tdt-0.6b-v3 (multilingual).
|
||||
|
||||
Modes:
|
||||
batch — transcribe the whole audio at once, print at the end (default).
|
||||
stream — sliding window with LocalAgreement: progressive text output as
|
||||
new audio arrives. Each window of WINDOW seconds is re-transcribed
|
||||
every SHIFT seconds; tokens agreed between two consecutive windows
|
||||
are "confirmed" and printed; the rest stays uncertain until the
|
||||
next iteration confirms (or invalidates) it.
|
||||
"""
|
||||
|
||||
import argparse
|
||||
import os
|
||||
import subprocess
|
||||
import sys
|
||||
import tempfile
|
||||
from pathlib import Path
|
||||
|
||||
MODEL_NAME = "nvidia/parakeet-tdt-0.6b-v3"
|
||||
|
||||
# Sliding-window streaming defaults (in seconds).
|
||||
# Matches whisper.cpp/examples/stream defaults.
|
||||
DEFAULT_WINDOW_S = 10.0
|
||||
DEFAULT_SHIFT_S = 3.0
|
||||
|
||||
|
||||
def download_audio(url: str, out_path: str):
|
||||
subprocess.run([
|
||||
|
|
@ -24,46 +38,142 @@ def convert_wav(src: str, dst: str):
|
|||
], check=True, capture_output=True)
|
||||
|
||||
|
||||
def transcribe(wav_path: str, long_audio: bool = True) -> str:
|
||||
def load_model(long_audio: bool = True):
|
||||
import nemo.collections.asr as nemo_asr
|
||||
import torch
|
||||
|
||||
print(f"Loading {MODEL_NAME} ...", flush=True)
|
||||
asr = nemo_asr.models.ASRModel.from_pretrained(MODEL_NAME)
|
||||
asr = asr.cuda() if __import__("torch").cuda.is_available() else asr
|
||||
|
||||
asr = asr.cuda() if torch.cuda.is_available() else asr
|
||||
if long_audio:
|
||||
asr.change_attention_model(
|
||||
self_attention_model="rel_pos_local_attn",
|
||||
att_context_size=[256, 256],
|
||||
)
|
||||
return asr
|
||||
|
||||
|
||||
def transcribe_batch(asr, wav_path: str) -> str:
|
||||
print(f"Transcribing {wav_path} ...", flush=True)
|
||||
out = asr.transcribe([wav_path], timestamps=False)
|
||||
return out[0].text
|
||||
|
||||
|
||||
def main():
|
||||
if len(sys.argv) < 2:
|
||||
print(f"Usage: {sys.argv[0]} <wav-file-or-youtube-url>")
|
||||
sys.exit(1)
|
||||
def _norm(w: str) -> str:
|
||||
"""Normalize a word for dedup comparison: lower-case, strip punctuation."""
|
||||
return "".join(ch for ch in w.lower() if ch.isalnum())
|
||||
|
||||
src = sys.argv[1]
|
||||
|
||||
def _dedup_overlap(prev_text: str, curr_text: str) -> str:
|
||||
"""Return the new portion of curr_text by removing the trailing words of
|
||||
prev_text that match the leading words of curr_text (overlap region).
|
||||
Comparison is case- and punctuation-insensitive."""
|
||||
prev_words = prev_text.split()
|
||||
curr_words = curr_text.split()
|
||||
prev_norm = [_norm(w) for w in prev_words]
|
||||
curr_norm = [_norm(w) for w in curr_words]
|
||||
max_n = min(len(prev_norm), len(curr_norm))
|
||||
for n in range(max_n, 0, -1):
|
||||
if prev_norm[-n:] == curr_norm[:n]:
|
||||
return " ".join(curr_words[n:])
|
||||
return curr_text
|
||||
|
||||
|
||||
def transcribe_sliding(asr, wav_path: str, window_s: float, shift_s: float) -> str:
|
||||
"""Sliding-window streaming: each window of WINDOW seconds is transcribed
|
||||
every SHIFT seconds; words overlapping with the previous window are
|
||||
de-duplicated and only the new tail is printed."""
|
||||
import soundfile as sf
|
||||
|
||||
if shift_s <= 0 or window_s <= 0:
|
||||
raise ValueError("window and shift must be positive")
|
||||
if shift_s > window_s:
|
||||
raise ValueError("shift must be <= window")
|
||||
|
||||
audio, sr = sf.read(wav_path)
|
||||
if sr != 16000:
|
||||
raise ValueError(f"Expected 16kHz WAV, got {sr}Hz")
|
||||
|
||||
n_samples = len(audio)
|
||||
win = int(window_s * sr)
|
||||
shift = int(shift_s * sr)
|
||||
overlap_s = window_s - shift_s
|
||||
|
||||
print(
|
||||
f"Sliding window: {window_s:.1f}s window, {shift_s:.1f}s shift "
|
||||
f"({overlap_s:.1f}s overlap), audio {n_samples / sr:.1f}s",
|
||||
flush=True,
|
||||
)
|
||||
print("--- streaming transcription ---", flush=True)
|
||||
|
||||
tmp = tempfile.NamedTemporaryFile(suffix=".wav", prefix="parakeet-slide-",
|
||||
delete=False)
|
||||
tmp_wav = tmp.name
|
||||
tmp.close()
|
||||
|
||||
full_text_parts = []
|
||||
prev_text = ""
|
||||
pos = 0
|
||||
|
||||
try:
|
||||
while pos < n_samples:
|
||||
end = min(pos + win, n_samples)
|
||||
chunk = audio[pos:end]
|
||||
sf.write(tmp_wav, chunk, sr)
|
||||
|
||||
out = asr.transcribe([tmp_wav], timestamps=False, verbose=False)
|
||||
current = out[0].text.strip()
|
||||
|
||||
new_part = _dedup_overlap(prev_text, current) if prev_text else current
|
||||
if new_part:
|
||||
print(new_part, end=" ", flush=True)
|
||||
full_text_parts.append(new_part)
|
||||
|
||||
prev_text = current
|
||||
pos += shift
|
||||
|
||||
print(flush=True)
|
||||
return " ".join(full_text_parts).strip()
|
||||
finally:
|
||||
if os.path.exists(tmp_wav):
|
||||
os.unlink(tmp_wav)
|
||||
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser(description=__doc__.splitlines()[0])
|
||||
parser.add_argument("source", help="WAV file path or YouTube URL")
|
||||
parser.add_argument("--stream", action="store_true",
|
||||
help="streaming mode (sliding window + LocalAgreement)")
|
||||
parser.add_argument("--window", type=float, default=DEFAULT_WINDOW_S,
|
||||
help=f"sliding window length in seconds (default: {DEFAULT_WINDOW_S})")
|
||||
parser.add_argument("--shift", type=float, default=DEFAULT_SHIFT_S,
|
||||
help=f"window shift in seconds (default: {DEFAULT_SHIFT_S}); "
|
||||
"overlap = window - shift")
|
||||
args = parser.parse_args()
|
||||
|
||||
src = args.source
|
||||
|
||||
if src.startswith(("http://", "https://", "youtu", "www.")):
|
||||
with tempfile.TemporaryDirectory(prefix="parakeet-") as tmp:
|
||||
raw = os.path.join(tmp, "audio.%(ext)s")
|
||||
wav = os.path.join(tmp, "audio_16k.wav")
|
||||
print(f"Downloading {src} ...", flush=True)
|
||||
download_audio(src, raw)
|
||||
downloaded = [f for f in os.listdir(tmp) if f.startswith("audio.")]
|
||||
print("Converting to 16kHz WAV ...", flush=True)
|
||||
convert_wav(os.path.join(tmp, downloaded[0]), wav)
|
||||
text = transcribe(wav)
|
||||
tmp = tempfile.mkdtemp(prefix="parakeet-")
|
||||
raw = os.path.join(tmp, "audio.%(ext)s")
|
||||
wav = os.path.join(tmp, "audio_16k.wav")
|
||||
print(f"Downloading {src} ...", flush=True)
|
||||
download_audio(src, raw)
|
||||
downloaded = [f for f in os.listdir(tmp) if f.startswith("audio.")]
|
||||
print("Converting to 16kHz WAV ...", flush=True)
|
||||
convert_wav(os.path.join(tmp, downloaded[0]), wav)
|
||||
wav_path = wav
|
||||
else:
|
||||
text = transcribe(src)
|
||||
wav_path = src
|
||||
|
||||
print("\n=== Transcription ===")
|
||||
print(text)
|
||||
asr = load_model(long_audio=not args.stream)
|
||||
|
||||
if args.stream:
|
||||
text = transcribe_sliding(asr, wav_path, args.window, args.shift)
|
||||
else:
|
||||
text = transcribe_batch(asr, wav_path)
|
||||
print("\n=== Transcription ===")
|
||||
print(text)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
|
|
|
|||
Loading…
Reference in New Issue