Fast Speech to Text — Parallel Whisper Transcription for Long Audio & Video

Transcribe long media fast by splitting it into windows and running many workers at once. Each run transcribes one window — or a whole list of windows with the model kept warm — using faster-whisper (large-v3-turbo) on CPU with word-level timestamps.

Use it for: hour-long podcasts, webinars, lecture archives, call recordings, or any pipeline where waiting for a single sequential transcription is too slow.

• Word timestamps and confidence, 90+ languages
• One window per run, a batch of windows per warm worker, or a queue-driven worker pool
• Tail overlap with _overlap flags so merged shards de-duplicate cleanly
• Range-reads the source when the server supports it — no need to host the whole file per shard
• Optional Standby HTTP mode for low-latency requests

Quick start — one window

{
  "source": "https://example.com/audio16k.mp3",
  "start_sec": 0,
  "duration_sec": 240,
  "overlap_sec": 1.5,
  "language": "en",
  "preset": "fast",
  "model": "large-v3-turbo"
}

Quick start — warm batch worker (recommended)

One run loads the model once and processes every window in the list:

{
  "source": "https://example.com/audio16k.mp3",
  "mode": "batch",
  "windows": [
    { "job_id": "j1", "shard_index": 0, "source": "https://example.com/audio16k.mp3", "start_sec": 0,   "duration_sec": 240, "overlap_sec": 1.5 },
    { "job_id": "j1", "shard_index": 1, "source": "https://example.com/audio16k.mp3", "start_sec": 240, "duration_sec": 240, "overlap_sec": 1.5 }
  ]
}

Start N runs like this in parallel and a 60-minute file finishes in minutes.

Input

Field	Default	Notes
source	—	required. https:// URL or kv:<storeId>/<key>
start_sec	0	offset of this window on the original timeline
duration_sec	240	window length. 240–300 s is the measured sweet spot
overlap_sec	1.5	extra tail audio; those words are flagged _overlap
language	auto	ISO code (en, vi, …)
preset	fast	fast | adaptive | balanced | accurate
model	large-v3-turbo	also distil-large-v3 (English only), medium
vad_filter	false	skip silence before decoding
cpu_threads	4	best value on a 16 GB run (≈4 vCPU)
batch_size	8	VAD chunks decoded in parallel; 0/1 = sequential
allow_empty	true	silent windows return 0 words instead of failing
mode	""	"" one window, batch window list, pool_worker queue
windows	—	window list for mode: "batch"
queue_id, dataset_id, worker_label	—	pool_worker wiring
idle_sec, max_life_sec, max_jobs	25, 600, 64	worker leashes

distil-large-v3 is English-only and is automatically downgraded to large-v3-turbo for non-English audio.

Output

One dataset row per window:

{
  "status": "success",
  "shard_index": 0,
  "start_sec": 0,
  "duration_sec": 240,
  "language": "en",
  "words": [
    { "text": "Hello", "startMs": 120, "endMs": 410, "confidence": 0.97 },
    { "text": "everyone", "startMs": 420, "endMs": 760, "confidence": 0.95, "_overlap": true }
  ],
  "segments": [{ "startMs": 120, "endMs": 4120, "text": "Hello everyone, welcome back" }],
  "meta": { "model": "large-v3-turbo", "rtf": 4.7, "decode_sec": 51.2 }
}

All timestamps are already offset back to the original timeline, so merging shards is: concatenate, drop words flagged _overlap that duplicate the next shard's first words, sort by startMs.

Worker pool mode

For very large fan-outs, put the windows in an Apify request queue and start N workers with mode: "pool_worker". fetch_next_request is atomic, so two workers never take the same window, and every worker pushes rows into one shared dataset.

{
  "source": "https://example.com/audio16k.mp3",
  "mode": "pool_worker",
  "queue_id": "<requestQueueId>",
  "dataset_id": "<datasetId>",
  "worker_label": "w1",
  "idle_sec": 25,
  "max_life_sec": 600,
  "max_jobs": 64
}

Queue mode requires the workers to run under the same account with full permissions.

Error handling

reason	Meaning	What to do
BAD_INPUT	missing source, invalid window	check the payload
UPSTREAM_BLOCKED	host refused the range request	host the audio yourself first
TIMEOUT	window too long for the run timeout	reduce duration_sec
OOM_LIMIT	not enough memory	run with 16 GB
INTERNAL	unexpected failure	retry that window only

Retries are cheap: a failed window is one shard, not the whole file.

Performance

16 GB run (≈4 vCPU), large-v3-turbo, preset: fast, batch_size: 8:

• real-time factor ≈ 4.7× (a 240 s window decodes in ≈50 s)
• 60-minute file, 15 parallel workers → ≈3–5 minutes wall clock
• windows shorter than ~60 s waste fixed overhead; 240–300 s is optimal

FAQ

One window or batch mode? Batch, whenever you have more than a couple of windows — the model is loaded once instead of per run.

How do I prepare the audio? Extract a mono 16 kHz track once (see the Video & Audio Toolkit actor, op: "audio_full") and point every shard at that one URL.

Do I need a GPU? No. CPU only.

Can I get speaker labels? Not here — pair the merged transcript with the Speaker Diarization actor.