Speech to Text: Whisper Audio & Video Transcriber

Transcribe audio and video to text with Whisper running inside Apify. Upload files or paste media URLs, then receive clean transcripts, timestamped segments, and ready-to-use SRT, VTT, TXT, or JSON output. Language detection, translation to English, word timestamps, silence skipping, and batch processing are built in—no external speech-to-text API key is required.

🎙️ What does Speech to Text: Whisper Audio & Video Transcriber do?

1. Reads each direct audio/video URL or uploaded file from your input, along with your model and language options.
2. Streams each download to disk with retries, size protection, and timeouts instead of loading the whole file into memory.
3. Uses ffmpeg to decode the media and prepare it for transcription. Video files work when they contain an audio track.
4. Runs faster-whisper locally inside the Actor with the model you choose: tiny, base, small, medium, or large-v3.
5. Detects the spoken language automatically, or uses the ISO language code you provide. It can transcribe the original language or translate the speech to English.
6. Optionally skips long silences with voice activity detection (VAD) and adds word-level timestamps to each segment.
7. Saves the selected transcript formats to the run's key-value store and adds SRT, VTT, and TXT links to the dataset item.
8. Processes every source independently. If one file is corrupt, has no audio, or cannot be downloaded, that item is marked failed; the rest of the batch continues.
9. Saves a run summary under the OUTPUT key, including item counts, audio processed, timing, and errors.

👥 Who is it for?

• Podcasters who need transcripts and subtitle files from episodes.
• Researchers working with interviews, oral histories, lectures, and field recordings.
• Journalists turning recordings into searchable notes and quotes for review.
• Marketers and content teams repurposing webinars, videos, and social clips.
• Accessibility teams producing captions and text alternatives.
• Developers and AI builders preparing audio for search, RAG pipelines, summaries, or LLM workflows.
• Video creators who want SRT or VTT captions without operating a separate speech API.

⚡ Why use this Actor?

• No separate speech-to-text API key. Whisper runs inside the Actor container.
• Batch processing. Submit multiple URLs or uploads in one run and get one dataset item per source.
• Useful output formats. Request JSON, TXT, SRT, and VTT files to match your downstream workflow.
• Model choice. Use tiny for quick drafts or large-v3 when accuracy matters more than speed.
• Language detection or fixed language. Let Whisper detect the language, or provide an ISO code for more predictable results.
• Translation to English. Set task to translate when you need English text from non-English audio.
• Silence handling. VAD is enabled by default and can reduce wasted transcription on long pauses.
• Word-level timestamps. Enable wordTimestamps when you need timing inside each segment.
• Fault isolation. One bad file is recorded as failed and does not stop the batch.
• Transparent billing. Pay per started audio minute for successful files, with a price tier for every model. Failed files are not charged per minute.
• Streamed downloads and retries. The Actor handles redirects, temporary network errors, and download limits.
• Video-page support via yt-dlp. Public YouTube and other supported video pages can be used in addition to direct media URLs.

📄 Sample output

The example below uses the public-domain JFK inaugural address sample distributed with whisper.cpp:

https://github.com/ggerganov/whisper.cpp/raw/master/samples/jfk.wav

Source	Language	Duration	Segments	Status	Files
jfk.wav	en	11.0 s	1	success	SRT, VTT, TXT

The transcript text and segment data below are illustrative; timing, confidence, punctuation, and example links vary by model, audio, and run.

{
  "sourceUrl": "https://github.com/ggerganov/whisper.cpp/raw/master/samples/jfk.wav",
  "text": "And so my fellow Americans, ask not what your country can do for you, ask what you can do for your country.",
  "language": "en",
  "languageProbability": 0.97,
  "duration": 11.0,
  "processingTime": 2.6,
  "segments": [
    {
      "start": 0.0,
      "end": 11.0,
      "text": "And so my fellow Americans, ask not what your country can do for you, ask what you can do for your country."
    }
  ],
  "status": "success",
  "error": null,
  "srtUrl": "https://api.apify.com/v2/key-value-stores/AbCdEf123456/records/transcripts-0001-jfk.wav.srt",
  "vttUrl": "https://api.apify.com/v2/key-value-stores/AbCdEf123456/records/transcripts-0001-jfk.wav.vtt",
  "txtUrl": "https://api.apify.com/v2/key-value-stores/AbCdEf123456/records/transcripts-0001-jfk.wav.txt"
}

When wordTimestamps is enabled, each segment also includes a words array with per-word timing and confidence.

🎛️ Input

At least one source is required: provide audioUrls, upload files through audioFiles, or use both.

{
  "audioUrls": [
    "https://github.com/ggerganov/whisper.cpp/raw/master/samples/jfk.wav"
  ],
  "modelSize": "base",
  "language": "auto",
  "task": "transcribe",
  "outputFormats": ["json", "txt", "srt", "vtt"],
  "wordTimestamps": false,
  "vadFilter": true
}

Field	Type	Default	Description
audioUrls	array of strings	[]	Direct audio or video URLs. Public YouTube and other supported video pages can be processed through yt-dlp. Direct media URLs are the most reliable option.
audioFiles	array of strings	[]	Files uploaded through the input's file upload field. They are stored in a key-value store and transcribed like URL sources.
modelSize	string enum	base	Whisper model: tiny, base, small, medium, or large-v3.
language	string	auto	auto for language detection, or an ISO language code such as en, de, es, or zh.
task	string enum	transcribe	transcribe keeps the original language; translate produces English text.
outputFormats	array of strings	json, txt, srt, vtt	Transcript files to save in the key-value store. The structured result is always returned in the dataset.
wordTimestamps	boolean	false	Adds word-level timing and confidence inside each segment.
vadFilter	boolean	true	Uses voice activity detection to skip silence.
maxConcurrency	integer	0 (automatic)	Controls how many files are transcribed concurrently. 0 lets the Actor choose a safe value for the allocated memory and CPU.
beamSize	integer	5	Beam search width. Higher values may improve recognition at the cost of speed.
temperature	number	0	Decoding temperature. 0 uses Whisper's standard fallback behavior; higher values allow more sampling.
initialPrompt	string	empty	Optional vocabulary or context hint for names, jargon, acronyms, or expected wording.

📦 Output

Each source produces one dataset item, including failed sources.

Field	Type	Description
sourceUrl	string / null	Resolved URL used for the source; may be empty for an uploaded file without an HTTP URL.
text	string	Full transcript. Empty for failed items and for audio where no speech is detected.
language	string / null	Detected or requested language code.
languageProbability	number / null	Confidence of the language detection.
duration	number / null	Audio duration in seconds.
processingTime	number	Wall-clock processing time for the item in seconds. This varies by model, hardware, and audio.
segments	array of objects	Timestamped segments, each with start, end, and text. When wordTimestamps is enabled, each segment also includes words.
status	string	success or failed.
error	string / null	Clear error message when status is failed; null on success.
srtUrl	string / null	Link to the SRT file when SRT output was requested and the item succeeded.
vttUrl	string / null	Link to the VTT file when VTT output was requested and the item succeeded.
txtUrl	string / null	Link to the TXT file when TXT output was requested and the item succeeded.

Run summary: the OUTPUT key in the run's key-value store contains the summary for the whole run, including processed and failed counts and the errors collected during the batch.

Downloading results

• Open the run in Apify Console and use the Output tab to view or export the dataset.
• Use the API with the run's dataset ID:

curl "https://api.apify.com/v2/datasets/DATASET_ID/items?format=json&clean=true" \
  -H "Authorization: Bearer $APIFY_TOKEN"

SRT, VTT, and TXT files are available through the links in each dataset item and in the run's key-value store.

🧠 Choosing a model

modelSize	Best for	Relative speed	Relative accuracy	Recommended Actor memory
tiny	Quick drafts, clear speech, low-resource runs	Fastest	Basic	2048 MB
base	Everyday transcription and the default balance	Fast	Good	4096 MB
small	Accents, noisier recordings, better detail	Moderate	Better	8192 MB
medium	High-stakes transcripts and complex audio	Slower	High	8192 MB
large-v3	Difficult audio where accuracy matters most	Slowest	Highest	16384 MB

These are relative descriptions, not benchmark guarantees. Actual speed depends on the allocated CPU and memory, the number and length of files, and the selected options. On Apify, more memory also increases the available CPU share.

The Actor minimum is 2048 MB. A 2 GB run with the base model was verified with a 380 MB peak memory footprint. Use 4 GB or more for faster CPU throughput on long recordings.

The Actor image includes the base model. Selecting another size can add startup time on its first use while that model is prepared.

💳 Pricing

This Actor uses pay-per-event billing. Prices depend on the selected model tier:

Event name	Price	When it applies
apify-actor-start	$0.00005 per memory unit (Apify default)	Charged automatically once per run. A 4 GB run is charged 4 units = $0.0002.
audio-minute-basic	$0.02 per started audio minute	Successful files transcribed with tiny or base. This is the primary event.
audio-minute-small	$0.04 per started audio minute	Successful files transcribed with small.
audio-minute-medium	$0.08 per started audio minute	Successful files transcribed with medium.
audio-minute-large	$0.16 per started audio minute	Successful files transcribed with large-v3.
apify-default-dataset-item	Removed	No per-item charge.

A run uses one model, so exactly one per-minute event is charged. Billing rounds up to the next whole minute, with a one-minute minimum for each successfully transcribed file. Failed files do not incur per-minute charges, but the automatic actor-start event still applies when a run starts. Platform usage is included in the per-minute prices shown above; no separate platform-usage charge is added for standard runs.

Audio length	Model	Event	Billable minutes	Per-minute cost
12:40	tiny / base	audio-minute-basic	13	$0.26
12:40	small	audio-minute-small	13	$0.52
12:40	medium	audio-minute-medium	13	$1.04
12:40	large-v3	audio-minute-large	13	$2.08
1:00:00	tiny / base	audio-minute-basic	60	$1.20
1:00:00	small	audio-minute-small	60	$2.40
1:00:00	medium	audio-minute-medium	60	$4.80
1:00:00	large-v3	audio-minute-large	60	$9.60
20:00, failed file	Any	Any	0	$0.00

At the default 4 GB Actor memory, add $0.0002 for the actor-start event to each run.

💡 Use cases

• Podcast and interview transcription for notes, show pages, and search.
• Subtitle generation for video libraries using SRT or VTT.
• Research archives that need searchable transcripts and timestamped quotes.
• Accessibility workflows for captions and text versions of spoken content.
• Content repurposing into articles, summaries, social posts, and newsletters.
• AI and RAG pipelines that need clean text plus segment timing.
• Call and meeting analysis where batch processing and per-file failures matter.

✨ Tips for best accuracy

• Use a larger model when the recording has accents, background noise, or specialized vocabulary.
• Set language explicitly when you already know the spoken language; automatic detection is convenient but less predictable on very short or noisy clips.
• Use initialPrompt for recurring names, product terms, or acronyms.
• Feed the cleanest audio you have. Excessive clipping, music, and overlapping speakers make transcription harder for any Whisper model.
• Keep vadFilter enabled for recordings with long pauses; disable it if quiet speech is being skipped.
• Enable wordTimestamps only when you need word timing; it adds processing time.
• If a long recording returns a short transcript, compare the audio duration with the last segment end. VAD may be excluding silence; set vadFilter: false to process the full timeline, but expect possible hallucinations over silence.
• Test a representative sample before processing a large batch.

📌 Limitations

• Whisper is not guaranteed to be perfect. Accents, technical terms, music, overlapping speech, and poor recordings can cause errors. Review important transcripts.
• No speaker labels or diarization. The output tells you what was said, not who said it.
• Translation is English-only. The translate task converts speech to English; it is not a general translation service.
• CPU inference takes time. Larger models and long recordings take longer than tiny or base.
• Public video pages only. YouTube and other video-page support depends on yt-dlp and public access. Login-gated, private, DRM-protected, or geo-restricted media may fail.
• No live transcription. This is a batch Actor, not a real-time streaming service.
• Word timestamps are estimates. They can be less precise on music-heavy or overlapping audio.
• Authentication-protected media URLs are not supported unless the source is directly accessible to the Actor.
• A silent recording may produce an empty transcript. It can still complete successfully because the file itself is valid.

❓ FAQ

How do I convert an MP3 to text?
Upload the MP3 through audioFiles or paste a direct URL into audioUrls, then run the Actor. The transcript is returned in the dataset.

Can I turn a YouTube video into text?
Yes, for public pages supported by yt-dlp. Direct media URLs are more reliable; private, age-restricted, or DRM-protected videos are not supported.

Which languages are supported?
Whisper is multilingual. Use auto for detection or provide an ISO language code. Translation, when enabled, is to English only.

Do I need an API key?
No separate speech-to-text API key is needed. You need an Apify account to run the Actor and an Apify API token only if you call it programmatically.

Can I generate SRT or VTT subtitles?
Yes. Include srt, vtt, or both in outputFormats. The dataset item links to the files saved in the key-value store.

Can I transcribe more than one file per run?
Yes. Add multiple URLs or uploads; each source becomes its own dataset item.

What happens if one file fails?
The item is returned with status: "failed" and an error message. Other files continue processing, and the failed file is not billed.

Can it identify different speakers?
No. The Actor does not provide speaker labels or diarization.

Can I get word-level timestamps?
Yes. Set wordTimestamps: true to add word timing and confidence inside each segment.

Can I call it from code or an AI agent?
Yes. Use the Apify API or an Apify client. It can also be run through Apify's MCP server in compatible MCP clients.

Is this real-time transcription?
No. It is designed for batch jobs and recordings, not live audio streams.

Which model should I start with?
Start with base. Move to small or larger when accuracy matters, or use tiny for quick drafts and very clear audio.

🛠️ Troubleshooting

What you may see	Likely cause	Fix
The URL returned text/html instead of audio	The link points to a web page, not a media file	Use a direct media URL, upload the file, or use a supported public video page.
No audio track was found in the file	The video or container has no audio stream	Use a source that contains audio.
ffprobe could not read the input	The file is corrupt, truncated, or uses an unsupported container	Re-encode it to a common format such as MP3, WAV, or M4A and retry.
The audio is empty or has no measurable duration	The file has no usable audio content	Check the recording with a media player or ffprobe.
HTTP 403, HTTP 404, or a download error	The URL is private, expired, moved, or blocked	Use a fresh direct URL or upload the file.
HTTP 429 or a timeout	The host is rate-limiting or slow	Retry later; the Actor already retries transient download failures.
The file is larger than the limit	The media exceeds the Actor's download limit	Use a smaller or lower-bitrate version, or split the recording.
Transcription failed: ...	The model or decoder could not process this audio	Retry once, try a smaller model, or clean up the audio.
Model loading failed	The selected model could not be prepared	Retry the run or choose a smaller model.
Empty text with status: success	The recording contains no detected speech, or VAD skipped it	Check the audio; try a clearer recording or disable VAD if quiet speech is present.

🧑‍💻 API usage

The examples use the deployed Actor ID lightning_ak/speech-to-text-whisper.

Python

from apify_client import ApifyClient

client = ApifyClient("YOUR_APIFY_TOKEN")

run = client.actor("lightning_ak/speech-to-text-whisper").call(
    run_input={
        "audioUrls": ["https://example.com/interview.mp3"],
        "modelSize": "base",
        "language": "auto",
        "task": "transcribe",
        "outputFormats": ["json", "srt", "vtt", "txt"],
        "wordTimestamps": True,
        "vadFilter": True,
    },
)

items = client.dataset(run["defaultDatasetId"]).list_items().items
print(items)

JavaScript

import { ApifyClient } from 'apify-client';

const client = new ApifyClient({ token: 'YOUR_APIFY_TOKEN' });

const run = await client.actor('lightning_ak/speech-to-text-whisper').call({
  audioUrls: ['https://example.com/interview.mp3'],
  modelSize: 'base',
  language: 'auto',
  task: 'transcribe',
  outputFormats: ['json', 'srt', 'vtt', 'txt'],
  wordTimestamps: true,
  vadFilter: true,
});

const { items } = await client.dataset(run.defaultDatasetId).listItems();
console.log(items);

cURL

curl -X POST "https://api.apify.com/v2/acts/lightning_ak~speech-to-text-whisper/runs?waitForFinish=300" \
  -H "Authorization: Bearer $APIFY_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "audioUrls": ["https://example.com/interview.mp3"],
    "modelSize": "base",
    "language": "auto",
    "task": "transcribe",
    "outputFormats": ["json", "srt", "vtt", "txt"],
    "wordTimestamps": true,
    "vadFilter": true
  }'

Read the dataset using the defaultDatasetId from the run response:

curl "https://api.apify.com/v2/datasets/DATASET_ID/items?format=json&clean=true" \
  -H "Authorization: Bearer $APIFY_TOKEN"

Apify MCP: this Actor can also be discovered and run through Apify's MCP server in compatible AI clients, making it possible to transcribe media from an agent workflow.

⚖️ Legal and responsible use

Only process audio and video that you have the right to use. You are responsible for obtaining consent from speakers, respecting copyright, and following the terms of the platform that hosts the source media. Do not upload confidential, sensitive, or personal recordings without authorization. Transcripts can contain errors and should not be treated as a substitute for professional legal, medical, or compliance review.

💬 Feedback and support

Found a bug or need a different output format? Open an issue from the Actor's page in Apify Console and include the input type (URL or upload), selected modelSize, and the error message. Do not post private audio, personal data, or API tokens.

🔗 Related Actors

Explore related tools in the Apify Store:

• AI Actors
• Videos Actors
• Automation Actors