Speech to Text (Whisper) — Video & Audio Transcription with Word Timestamps

Turn any audio or video file into an accurate transcript with word-level timestamps, confidence scores and ready-to-use caption segments. Runs whisper.cpp on CPU — no GPU, no API keys, no local install.

Use it for: subtitles and SRT-style captions, searchable video archives, podcast show notes, meeting notes, content repurposing, dataset labelling.

• 90+ languages, or automatic language detection
• Word timestamps, so captions can be split to any length
• Accepts MP4, MOV, MKV, WebM, MP3, WAV, M4A, OGG… (converted internally to 16 kHz mono)
• Accepts a public URL or a key-value-store record from a previous run
• Optional time window, so long media can be sharded across parallel runs

Quick start

{
  "op": "transcribe",
  "source": "https://example.com/interview.mp4",
  "options": { "preset": "balanced", "language": "en" }
}

Feature-detect the build (free, a few seconds, needs no source):

{ "op": "capabilities" }

Input

Field	Type	Notes
op	transcribe | capabilities	default transcribe
source	string	https:// URL or kv:<storeId>/<key>. Audio or video.
options.preset	fast | adaptive | balanced | accurate	speed vs accuracy
options.language	ISO code (en, vi, es…)	omit for auto-detect
options.model	model name	overrides the preset default
options.threads	integer	defaults to the run's vCPU count
options.segment_gap_ms	number	silence gap that starts a new caption segment
options.start_sec / duration_sec / overlap_sec	number	transcribe one window only
options.keep_raw_json	boolean	also store the raw whisper output
output.signed_upload_url	string	PUT transcript.json into your own storage
cleanup	on_success | always | off	artifact retention
callback	object	{ url, secret_header: { name, value } } webhook

Presets

Preset	Best for	Relative speed
fast	drafts, search indexing	fastest
adaptive	mixed content, unknown quality	fast
balanced	default choice for captions	medium
accurate	publishing, noisy audio, accents	slowest

Output

{
  "status": "success",
  "op": "transcribe",
  "artifacts": [{ "name": "transcript", "kv_key": "transcript.json", "url": "https://api.apify.com/v2/key-value-stores/.../transcript.json" }],
  "meta": {
    "language": "en",
    "duration_sec": 338.4,
    "word_count": 912,
    "model": "…",
    "words": [
      { "id": 0, "text": "Hello", "startMs": 120, "endMs": 410, "confidence": 0.98 }
    ],
    "segments": [
      { "startMs": 120, "endMs": 4120, "text": "Hello and welcome back to the show" }
    ]
  }
}

• words[] — one entry per word with start/end in milliseconds and confidence.
• segments[] — sentence-like caption blocks, split on pauses (segment_gap_ms).
• The full transcript is also stored as the transcript.json artifact.

Sharding a long file

Pass a window and the actor transcribes only that slice; timestamps are offset back to the original timeline, and words inside the tail overlap are flagged _overlap so you can de-duplicate when merging shards.

{
  "op": "transcribe",
  "source": "https://example.com/audio16k.wav",
  "options": { "start_sec": 600, "duration_sec": 300, "overlap_sec": 1.5 }
}

For heavy parallel fan-out, use the dedicated Fast Speech to Text actor, which keeps the model warm and processes a list of windows per run.

Error handling

{ "ok": false, "reason": "BAD_INPUT", "message": "..." }

reason	Meaning	What to do
BAD_INPUT	missing/unreadable source	verify the URL is publicly reachable
UPSTREAM_BLOCKED	the host refused the download	host the file yourself, or fetch it first
TIMEOUT	run exceeded its timeout	shard the media, or pick a faster preset
OOM_LIMIT	not enough memory	run with 16 GB
INTERNAL	unexpected failure	retry; report the run ID

Performance

Measured on a 16 GB Apify run (≈4 vCPU), CPU only:

Media length	fast	balanced	accurate
5 min	~1 min	~2 min	~4 min
30 min	~5 min	~10 min	~20 min
60 min	shard it	shard it	shard it

For anything over ~30 minutes, extract 16 kHz mono audio first (see the Video & Audio Toolkit actor) and transcribe in parallel windows.

FAQ

Does it need a GPU or an API key? No. It runs whisper.cpp on CPU inside the actor.

Which languages are supported? All Whisper languages (90+). Leave language empty to auto-detect, or set it explicitly for better accuracy.

Can I get SRT/VTT? The output gives word and segment timings in milliseconds; building SRT/VTT from that is a few lines of code, and lets you choose your own caption length.

Can I transcribe a YouTube link directly? Not here — resolve/download the media first with the Video Downloader or Video & Audio Toolkit actor, then pass the resulting URL as source.

Are speaker names included? No. Pair it with the Speaker Diarization actor, which can label this transcript with who spoke each word.