In this tutorial, we use Whistle, Cactus Compute’s speech-to-text model, a 16.9 MB file that runs on a CPU, transcribes seven languages, and returns word timestamps and probabilities. The model card claims it beats Whisper base on LibriSpeech; we test it on seeded test-clean and test-other samples, comparing Whistle with Whisper tiny and base on identical clips, judged by paired speaker-cluster intervals with Holm correction. We find a lead over greedy Whisper base on test-other, no significant gap against beam search, and a speed advantage that narrows as clips grow. Beyond the headline, we probe a 30-second window where the pinned engine silently drops the middle of longer audio, plus streaming latency, timestamps, keyword biasing, other languages, noise, embeddings, and speech-to-tool calls. Our figures are one-run samples from an Apple M4, and published full-split numbers are quoted, not measured.
import os
import io
import re
import sys
import json
import math
import time
import random
import hashlib
import shutil
import zipfile
import platform
import traceback
import subprocess
import collections
import urllib.error
import urllib.request
RESULTS = {}
VERDICTS = []
def banner(title):
print("\n" + "=" * 78)
print(title)
print("=" * 78)
def section(name):
def wrap(fn):
def run(*a, **kw):
banner(name)
t0 = time.time()
try:
out = fn(*a, **kw)
RESULTS[name] = out if isinstance(out, str) else "ok"
return out
except Exception as e:
RESULTS[name] = f"SKIPPED / FAILED -> {type(e).__name__}: {e}"
print(f"\n[!] {name} did not complete: {type(e).__name__}: {e}")
traceback.print_exc(limit=3)
return None
finally:
print(f"\n[{time.time() - t0:.0f} s]")
json.dump({"results": RESULTS, "verdicts": VERDICTS}, open("results.json", "w"), indent=1)
return run
return wrap
banner("1. Install Whistle, pin its weights and engine by Hub revision, transcribe one clip")
subprocess.run([sys.executable, "-m", "pip", "install", "-q", "cactus-needle==3.2.0", "openai-whisper==20250625",
"jiwer==4.0.0", "soundfile==0.14.0", "soxr==1.1.0"], check=True)
os.environ["NEEDLE_TELEMETRY"] = "0" # the package sends anonymous usage counts unless told not to
os.environ["HF_HUB_DISABLE_PROGRESS_BARS"] = "1"
import numpy as np
import soundfile as sf
import soxr
from huggingface_hub import hf_hub_download
# cactus-needle fetches three files on first use, from Hugging Face repos whose files are replaced in place: the 16.9 MB
# whistle.cact weights, the ~1 MB engine library (version 3.3.0 in this package; the needle3 repo received new binaries
# on 2026-10-10) and the 35 MB needle3.cact used in step 12. Pinning the Hub revision of each, and printing the hashes,
# is what makes this run reproducible. The engine version is read from the package, not typed in.
WHISTLE_REV = "ca5287601bef25af26dcf2e1b2bdc0843a7c19e5" # Cactus-Compute/whistle, 2026-10-10
NEEDLE3_REV = "93e88ede867e4311c39151285e74df513e4b835b" # Cactus-Compute/needle3 (engine + Needle 3 weights), 2026-10-10
from needle.agent import fetch as needle_fetch
TAG = needle_fetch._platform_tag()
ENGINE_VERSION = needle_fetch.engine_version(3)
weights = hf_hub_download("Cactus-Compute/whistle", "whistle.cact", revision=WHISTLE_REV)
wheel = hf_hub_download("Cactus-Compute/needle3", needle_fetch.engine_wheel(ENGINE_VERSION, TAG, 3), revision=NEEDLE3_REV)
with zipfile.ZipFile(wheel) as z:
member = next(n for n in z.namelist() if n.startswith("needle/libneedle"))
os.makedirs("engine", exist_ok=True)
engine = os.path.abspath(os.path.join("engine", os.path.basename(member)))
open(engine, "wb").write(z.read(member))
needle3_weights = hf_hub_download("Cactus-Compute/needle3", "needle3.cact", revision=NEEDLE3_REV)
os.makedirs(needle_fetch.cache_dir(3), exist_ok=True) # Needle(tools=...) reads its weights from this cache path
shutil.copyfile(needle3_weights, os.path.join(needle_fetch.cache_dir(3), needle_fetch.base_weights(3)))
os.environ["NEEDLE_WHISTLE_WEIGHTS"] = weights # both variables are honoured by the package
os.environ["NEEDLE3_LIB_PATH"] = engine
import needle
import importlib.metadata as md
sha = lambda p: hashlib.sha256(open(p, "rb").read()).hexdigest()[:12]
CPU = platform.processor() or platform.machine()
if sys.platform == "linux" and os.path.exists("/proc/cpuinfo"):
CPU = next((l.split(":", 1)[1].strip() for l in open("/proc/cpuinfo") if l.startswith("model name")), CPU)
elif sys.platform == "darwin":
CPU = subprocess.run(["sysctl", "-n", "machdep.cpu.brand_string"], capture_output=True, text=True).stdout.strip() or CPU
print(f"cactus-needle {needle.__version__} (engine {ENGINE_VERSION}, wheel tag {TAG}) python {platform.python_version()} "
f"soundfile {md.version('soundfile')} soxr {md.version('soxr')} numpy {np.__version__}")
print(f"hardware: {CPU}, {os.cpu_count()} cpus visible")
print(f"whistle.cact {os.path.getsize(weights)/1e6:.1f} MB sha256 {sha(weights)} @ {WHISTLE_REV[:12]}")
print(f"engine {os.path.getsize(engine)/1e6:.1f} MB sha256 {sha(engine)} @ {NEEDLE3_REV[:12]}")
print(f"needle3.cact {os.path.getsize(needle3_weights)/1e6:.1f} MB sha256 {sha(needle3_weights)} @ {NEEDLE3_REV[:12]}")
SR = 16000
W = needle.Whistle() # one model per process, not thread-safe; every later step reuses this object
def fetch(url, tries=7):
"""GET with retries. The datasets-server rate-limits anonymous callers (HTTP 429): honour Retry-After when it is
sent, otherwise back off 5, 10, 20, 40 ... seconds, and never give up on the first refusal."""
for k in range(tries):
try:
with urllib.request.urlopen(url, timeout=60) as r:
return r.read()
except urllib.error.HTTPError as e:
if k == tries - 1:
raise
wait = int(e.headers.get("Retry-After", 0) or 0) if e.code == 429 else 0
time.sleep(max(wait, 5 * 2 ** k) if e.code in (429, 500, 502, 503, 504) else 2)
except Exception:
if k == tries - 1:
raise
time.sleep(3 * (k + 1))
def page(dataset, config, split, offset, length=100):
"""One page of a Hub dataset through the datasets-server rows API: metadata plus signed audio URLs."""
url = (f"https://datasets-server.huggingface.co/rows?dataset={dataset}&config={config}"
f"&split={split}&offset={offset}&length={length}")
time.sleep(0.25) # stay under the anonymous rate limit when paging a whole split
return json.loads(fetch(url))
def load_audio(row):
"""Download one row's clip; return 16 kHz mono float32 samples and the sha256 of the original file."""
data = fetch(row["audio"][0]["src"])
audio, sr = sf.read(io.BytesIO(data), dtype="float32")
if audio.ndim > 1:
audio = audio.mean(axis=1)
if sr != SR:
audio = soxr.resample(audio, sr, SR, quality="HQ")
return audio, hashlib.sha256(data).hexdigest()[:12]
first = page("openslr/librispeech_asr", "clean", "test", 0, 1)["rows"][0]["row"]
clip, _ = load_audio(first)
t0 = time.perf_counter()
out = W.transcribe(clip, word_timestamps=True)
wall = (time.perf_counter() - t0) * 1000
print(f"\nclip {first['id']} {len(clip)/SR:.1f} s reference: {first['text']}")
print(f"-> {out['text']!r}")
print(f" language {out['language']} engine time to first token {out['ttft_ms']:.1f} ms decode {out['decode_tps']:.0f} tok/s"
f" wall {wall:.0f} ms")
print(" first words:", [(w["word"], w["start"], w["end"], w["probability"]) for w in out["words"][:3]])
RESULTS["1. Install Whistle, pin its weights and engine by Hub revision, transcribe one clip"] = \
f"{os.path.getsize(weights)/1e6:.1f} MB weights, {len(clip)/SR:.1f} s clip in {wall:.0f} ms"
We install cactus-needle 3.2.0 with its dependencies pinned and turn off its anonymous telemetry. The package downloads Whistle’s weights, its engine and Needle 3’s weights from Hugging Face repositories whose files are replaced in place, so we pin both repositories to fixed revisions and print SHA-256 hashes. The page() helper reads the datasets-server rows API with rate-limit backoff, and load_audio() resamples clips to 16 kHz. The first test-clean clip, 3.5 s of speech, transcribes in 33 ms of wall time, each word carrying its start, end and probability.
import jiwer
import whisper
import torch
from whisper.normalizers import EnglishTextNormalizer, BasicTextNormalizer
NORM_EN = EnglishTextNormalizer() # the normalisers Whistle's own benchmarks use; applied to both sides, always
NORM_BASIC = BasicTextNormalizer()
SEED = 2026
N_WHISTLE = 300 # clips per LibriSpeech split that Whistle transcribes (it is cheap)
N_PAIRED = {"other": 100, "clean": 60} # the first N of them also go through Whisper (it is not); test-other is primary
MAX_CLIP_S = 30.0 # Whistle's single-pass window; longer clips are declared out of scope here
WHISPER_SECONDS = collections.Counter()
def all_rows(dataset, config, split):
out, offset = [], 0
while True:
rows = page(dataset, config, split, offset)["rows"]
out += [r["row"] for r in rows]
offset += 100
if len(rows) < 100:
return out
def dataset_sha(dataset):
return json.loads(fetch(f"https://huggingface.co/api/datasets/{dataset}"))["sha"][:12]
def wer_counts(ref, hyp):
"""Errors (S+D+I) and reference words for one normalised pair; an empty hypothesis is all deletions."""
n = len(ref.split())
if not hyp.strip():
return n, n
o = jiwer.process_words(ref, hyp)
return o.substitutions + o.deletions + o.insertions, n
def sdi(ref, hyp):
if not hyp.strip():
return 0, len(ref.split()), 0
o = jiwer.process_words(ref, hyp)
return o.substitutions, o.deletions, o.insertions
def corpus_wer(errs, words):
return 100.0 * sum(errs) / max(sum(words), 1)
def _by_cluster(clusters, *arrays):
tot = collections.defaultdict(lambda: [0.0] * len(arrays))
for i, c in enumerate(clusters):
for j, a in enumerate(arrays):
tot[c][j] += a[i]
return tot
def wer_ci(clusters, errs, words, B=2000, seed=0):
"""95% percentile interval of a corpus WER, resampling clusters (speakers) with replacement."""
tot = _by_cluster(clusters, errs, words)
keys, rng = sorted(tot), random.Random(seed)
boots = []
for _ in range(B):
d = rng.choices(keys, k=len(keys))
boots.append(100 * sum(tot[c][0] for c in d) / max(sum(tot[c][1] for c in d), 1))
return np.percentile(boots, [2.5, 97.5])
def paired_contrast(clusters, errs_a, words, errs_b, label_a, label_b, B=2000, seed=0):
"""WER(a) - WER(b) with a paired speaker-cluster bootstrap: the same resampled speakers for both systems."""
tot = _by_cluster(clusters, errs_a, errs_b, words)
keys, rng = sorted(tot), random.Random(seed)
boots = []
for _ in range(B):
d = rng.choices(keys, k=len(keys))
w = max(sum(tot[c][2] for c in d), 1)
boots.append(100 * (sum(tot[c][0] for c in d) - sum(tot[c][1] for c in d)) / w)
boots = np.array(boots)
lo, hi = np.percentile(boots, [2.5, 97.5])
p = 2 * min((boots <= 0).mean(), (boots >= 0).mean())
return {"a": label_a, "b": label_b, "diff": corpus_wer(errs_a, words) - corpus_wer(errs_b, words),
"lo": lo, "hi": hi, "p": max(p, 1 / B), "clusters": len(keys)}
def holm(ps, alpha=0.05):
order, keep = np.argsort(ps), [False] * len(ps)
for rank, i in enumerate(order):
if ps[i] <= alpha / (len(ps) - rank):
keep[i] = True
else:
break
return keep
def verdict(c, significant):
"""Verdict rule: diff = WER(other system) - WER(Whistle). CI entirely > 0 and Holm-significant -> Whistle ahead;
entirely < 0 and significant -> Whistle behind; otherwise no significant difference at this sample size."""
if c["lo"] > 0 and significant:
return "Whistle ahead"
if c["hi"] < 0 and significant:
return "Whistle behind"
return "no significant difference at this n"
def fmt_contrast(c):
return f"{c['diff']:+.2f} pts (95% CI {c['lo']:+.2f}..{c['hi']:+.2f}, {c['clusters']} speakers)"
@section("2. Fixed, seeded samples of LibriSpeech test-clean and test-other, and the rules for the verdicts")
def sample_english():
global EN
print("Rules for the verdicts below, written into this notebook before its final run (exploratory probes on the")
print("same splits came first; the prose says which):")
print(f" * simple random sample of {N_WHISTLE} rows per split (seed {SEED}); the first {N_PAIRED['other']} (test-other) and "
f"{N_PAIRED['clean']} (test-clean) are the paired Whisper subsets")
print(f" * clips longer than {MAX_CLIP_S:.0f} s are excluded and counted (Whistle's single-pass window), as are empty references")
print(" * the same normaliser on both sides of each comparison (Whisper's EnglishTextNormalizer for LibriSpeech, its")
print(" BasicTextNormalizer for MLS); corpus WER = summed errors / summed reference words")
print(" * interval: 95% speaker-cluster paired bootstrap, 2,000 draws, the same resampled speakers for every system")
print(" * verdict per contrast: unadjusted CI of WER(other) - WER(Whistle) entirely > 0 and Holm-adjusted p < 0.05 ->")
print(" Whistle ahead; entirely < 0 and significant -> behind; otherwise no significant difference at this n")
print(" family A (published claims): test-other and test-clean vs base greedy and vs base beam 5, MLS pooled vs base greedy;")
print(" family B: the keyword contrasts. Holm at alpha = 0.05 within each family. Whisper tiny rows are descriptive.\n")
global MANIFEST
EN, manifest = {}, {"dataset_sha": dataset_sha("openslr/librispeech_asr"), "seed": SEED, "splits": {}}
MANIFEST = manifest
for split in ("other", "clean"):
t0 = time.time()
rows = all_rows("openslr/librispeech_asr", split, "test")
idx = random.Random(SEED).sample(range(len(rows)), N_WHISTLE)
kept, too_long, empty = [], 0, 0
for i in idx:
r = rows[i]
audio, h = load_audio(r)
ref = NORM_EN(r["text"])
if len(audio) / SR > MAX_CLIP_S:
too_long += 1
continue
if not ref:
empty += 1
continue
kept.append({"id": r["id"], "spk": str(r["speaker_id"]), "text": r["text"], "ref": ref,
"audio": audio, "dur": len(audio) / SR, "sha": h})
EN[split] = kept
manifest["splits"][split] = {"rows": len(rows), "sampled": N_WHISTLE, "too_long": too_long, "empty_ref": empty,
"ids": [c["id"] for c in kept], "sha": [c["sha"] for c in kept]}
durs = [c["dur"] for c in kept]
pk = kept[:N_PAIRED[split]]
print(f"test-{split}: {len(rows)} rows -> {len(kept)} clips kept ({too_long} over {MAX_CLIP_S:.0f} s, {empty} empty), "
f"{len(set(c['spk'] for c in kept))} speakers, {sum(durs)/60:.1f} min, clip {np.mean(durs):.1f} s mean / {max(durs):.1f} s max"
f" [{time.time()-t0:.0f} s]")
print(f" paired subset: {len(pk)} clips, {len(set(c['spk'] for c in pk))} speakers, {sum(c['dur'] for c in pk):.0f} s of audio, "
f"{sum(len(c['ref'].split()) for c in pk)} reference words")
json.dump(manifest, open("manifest.json", "w"))
print(f"\nmanifest.json: LibriSpeech revision {manifest['dataset_sha']}, clip ids and file hashes (the MLS and MInDS-14 clips are added to it in")
print("steps 9 and 12), so a rerun can prove it used the same audio")
return f"{len(EN['other'])} + {len(EN['clean'])} clips, manifest written"
sample_english()
A verdict means little if its rule is set after the final numbers arrive, so we wrote ours into the notebook before the final run, after exploratory probes that touched some of the same clips. We draw a seeded sample of 300 rows per split and drop clips longer than Whistle’s single-pass window, leaving 299 clips in each. The paired subsets, the first 100 test-other and 60 test-clean clips, go through Whisper, and few speakers make their intervals approximate. A difference counts only when its 95% paired speaker-cluster bootstrap interval excludes zero and its Holm-adjusted p is below 0.05.
WHISPER = {}
def whisper_model(size):
if size not in WHISPER:
WHISPER[size] = whisper.load_model(size, device="cpu")
return WHISPER[size]
def run_whisper(size, audio, language="en", **kw):
"""openai-whisper with its defaults (greedy with temperature fallback) unless kw says otherwise. Seeded per call,
because the fallback samples. Returns text, whether any segment needed the fallback, and the wall time."""
torch.manual_seed(0)
WHISPER_SECONDS[size] += len(audio) / SR
t0 = time.perf_counter()
r = whisper_model(size).transcribe(audio, language=language, fp16=False, **kw)
fell_back = any(s.get("temperature", 0) > 0 for s in r["segments"])
return r["text"], fell_back, time.perf_counter() - t0
SYSTEMS = {"other": [("whisper tiny, greedy", "tiny", {}), ("