Upload video
Try a real recording
- An interview
- Webinar
- YouTube-style video
- Product demonstration
- Meeting recording
- Lecture
- Podcast video
- Panel discussion
- Training session
Horizon · Video Transcription
Turn video speech into accurate, searchable text for subtitles, captions, interviews, meetings, podcasts, webinars, media archives, education, and AI workflows. Upload a video and use Shunya Horizon to extract and transcribe its audio across 204+ languages - with speaker labels, word-level timestamps, multilingual recognition, and optional conversation intelligence.
Welcome back to the show. Let's start with what changed this year.
Video → audio · HorizonVideo transcription
A video contains more information than its visuals. It contains interviews, conversations, presentations, narration, product discussions, lectures, customer stories, and sometimes hours of spoken information that cannot be searched until it becomes text. Horizon converts the audio track inside a video into structured transcripts.
204+
Languages supported across Horizon
3.10%
Composite WER across 8 OpenASR benchmarks
4.99%
Published WER on GigaSpeech - a benchmark including YouTube-style speech
5
Video input formats supported directly
| Capability | Horizon · Video Transcription |
|---|---|
| Video Transcription | |
| Automatic Audio Extraction | |
| MP4 Input | |
| MKV Input | |
| MOV Input | |
| AVI Input | |
| WebM Input | |
| Speaker Diarization | |
| Word-level Timestamps | |
| Confidence Scores | |
| Multilingual Transcription | |
| Code-switching | |
| Summarisation | |
| API Integration |
Try it yourself
Don't evaluate video transcription using a perfectly recorded studio clip.
Upload video
Supported formats
Language: Auto Detect - or choose the spoken language explicitly.
Model
Choose the Horizon model that matches the speech in your video.
Output
Video in, audio out
Video transcription does not mean that the speech-recognition model interprets every visual element onscreen. Shunya's batch workflow extracts the audio track from the uploaded video.
Batch workflow
Horizon answers
It does not, by itself, answer
For visual scene understanding, OCR, object recognition, slides, charts, or screen analysis, an additional video-vision layer would be required.
Important distinction: Horizon transcribes speech from the audio track, not on-screen visuals.
Video transcription accuracy
Videos do not all sound alike. A professionally recorded presentation is different from a podcast. A podcast is differentfrom a meeting. A meeting is different from an interview recorded on a phone. Shunya's OpenASRevaluation includes benchmarks representing several of these speech environments.
| Benchmark | Domain | Horizon WER |
|---|---|---|
| LibriSpeech Clean | Clean read speech | 0.71% |
| TED-LIUM | TED-style presentations | 1.43% |
| AMI | Multi-party meetings | 4.19% |
| GigaSpeech | Podcasts / YouTube | 4.99% |
| Earnings22 | Noisy earnings calls | 5.83% |
3.10%
Composite WER across all 8 evaluated benchmarks
Why video audio varies
Read this fairly
Benchmark results do not guarantee identical accuracy on every uploaded video. A clean studio interview and a street-recorded video can produce very different transcription conditions. The benchmark shows model performance. Your own videos provide the final test.
Who said what
Interviews, panels, meetings, podcasts, and discussions often contain multiple people. A plain transcript can lose the conversational structure.
Without diarization
Welcome back to the show. Thanks for having me. Let's start with what changed this year.
With diarization
SPEAKER_00 Welcome back to the show.
SPEAKER_01 Thanks for having me.
SPEAKER_00 Let's start with what changed this year.
Speaker diarization
Speaker identification
Diarization separates voices. Speaker identification can help resolve registered speakers into known identities.
Before
We launched the first version in March.
After
We launched the first version in March.
Workflow
Unregistered voices can remain anonymous speaker labels.
Word-level timestamps
A transcript becomes substantially more useful when every word can be linked back to the videotimeline. Horizon verbose output can include start time, end time, and confidence score.
Search
The transcript becomes an index into the video.
Subtitles & captions
Video transcription is often the first step toward captions and subtitles. Horizon can return word and segment timing information.That structured timing can then be converted by your application into subtitle formats such as SRT and WebVTT.
Subtitle format
Widely used for video subtitles.
Subtitle format
Commonly used for web-based video captions.
Caption workflow
Horizon provides the transcript and timing data required for subtitle generation. Your application can serialize those timestamps into the subtitle format required by the video platform.
Search inside video
A two-hour video may contain one sentence a user needs. Without transcription, recovery is slow. With transcription, search connects queries to timestamps in the original video.
Without transcription
With transcription
Index the transcript, run a query, jump to the matching moment.
Search workflow
Video knowledge
Multilingual video
Videos can include multilingual speakers, foreign-language interviews, English technical terminology, regional-language conversations,and code-switched discussions. Horizon provides multilingual speech recognition across 204+ languages.
204+
Languages supported
Model option
For broad global-language video transcription.
Model option
For videos primarily containing Indian-language speech.
Model option
For conversations where speakers naturally switch languages.
आज हम customer acquisition strategy के बारे में बात करेंगे.
Instead of forcing the complete sentence into Hindi or English, a code-switching workflow can preserve the way it was actually spoken.
One video
Video summaries
A transcript preserves everything that was said. A summary provides a faster view of the conversation.For supported batch workflows, Shunya's intelligence layer includes summarisation.
Video
From transcript to insight
From video to intelligence
A video may contain useful information beyond the literal spoken words. Shunya's optionalbatch intelligence layer can enrich completed transcription workflows.
Who spoke when?
Which registered speaker was talking?
What objective appears in the conversation?
What is the overall conversational sentiment?
What dominant emotion appears in different segments?
What happened across the recording?
Standardise important domain vocabulary.
Where exactly was each word spoken?
Mask organisation-defined sensitive terminology.
Mask profane words in transcript output.
Enrichment pipeline
Video to content
Once spoken video content becomes text, it can be reused by downstream applications.
Downstream assets
Important distinction
Horizon provides the speech transcript and optional intelligence. Content rewriting, article generation, chapter creation, action-item extraction, or social-content creation can then be handled by an LLM or downstream application.
Speech first
Generation second: transform it into the format you need.
Real-world use cases
Create searchable transcripts, captions, summaries, and downstream content workflows.
Separate interviewer and guest while preserving timestamps.
Turn long presentations and Q&A sessions into searchable written records.
Transcribe lectures, tutorials, lessons, and recorded classroom material.
Convert recorded video conferences into speaker-aware transcripts.
Transcribe video-podcast episodes for search, captions, notes, and archives.
Create transcript and subtitle workflows for production and post-processing.
Convert recorded interviews, user studies, and qualitative research sessions into text.
Make internal video libraries searchable and easier to review.
Transcribe customer interviews and testimonial recordings for internal and publishing workflows.
Media & post-production
Video editors frequently need to work with speech before the final video is complete.
Rough cut
Word timestamps make the transcript useful beyond simple reading.
Developers
Shunya's batch input supports both audio and video files. Submit a video file to POST /v1/audio/transcriptions. Choose Horizon Universal (horizon-universal), Horizon Indic (horizon-indic), or the Horizon model that fits the audio. The service extracts the audio track, processes speech, and returns text, segments, detected language, and timing information.
{
"model": "horizon-universal",
"file": "interview.mp4",
"enable_diarization": true,
"response_format": "verbose_json"
}Video formats
Optional
What happens
Basic workflow
Choose the right model
An MP4 could contain English, Hindi, Tamil, or ten speakers moving betweenlanguages. Choose the model according to what is being spoken.
Global video
Best for English, international interviews, global webinars, and multilingual media.
Best for:
Indian-language video
Best for Hindi, Tamil, Telugu, Bengali, Marathi, and other supported Indian languages.
Best for:
Mixed-language video
Best for Hinglish, multilingual interviews, Indian-language + English discussions, and code-switched podcasts.
Best for:
Video format
The video format, such as MP4, does not determine the speech model. The language and the conversation do.
Built for production
Uploading an MP4 and displaying a paragraph of text is only the beginning. Production video transcriptionneeds format support, multilingual recognition, speaker structure, accurate timing, APIs, intelligence,private deployment, and connections to the systems that use the transcript.
| Requirement | Horizon |
|---|---|
| Video Transcription | |
| MP4 | |
| MKV | |
| MOV | |
| AVI | |
| WebM | |
| Automatic Audio Extraction | |
| 204+ Languages | |
| Speaker Diarization | |
| Speaker Identification | |
| Word-level Timestamps | |
| Caption-ready Timing Data | |
| Multilingual Recognition | |
| Code-switching | |
| Summarisation | |
| Intent / Sentiment / Emotion | |
| Redaction | |
| API Integration | |
| Enterprise Deployment |
Production path
Questions
Video transcription converts the spoken audio inside a video into written text.
Yes. Shunya's batch transcription input accepts video files and extracts their audio track for speech recognition.
Shunya batch transcription accepts MP4, MKV, MOV, AVI, and WebM.
Horizon transcribes the extracted audio track. It does not by itself describe scenes, recognise visual objects, read text displayed on screen, or interpret charts and slides.
Horizon currently supports speech recognition across 204+ languages.
Yes. Speaker diarization can separate different voices into speaker-labelled transcript segments.
Yes. Speaker identification can resolve registered voices into known identities where configured.
Yes. Verbose transcription output can provide word-level timing and confidence information.
Yes. Word and segment timestamps provide the structured information needed to generate subtitle files such as SRT or WebVTT.
The batch workflow provides transcript timing information that can be serialised by your application into SRT or VTT.
Horizon's published benchmark suite includes GigaSpeech, representing podcast and YouTube-style speech, where Shunya reports 4.99% WER. That benchmark should not be treated as a guarantee for every online video.
Yes. Video interviews can be transcribed with speaker diarization and word-level timestamps.
Yes. Recorded presentations, webinars, lectures, and similar spoken video content can be processed through the batch transcription workflow.
For supported batch workflows, the transcript can be enriched using Shunya's summarisation capability.
Yes. Horizon supports multilingual transcription, and the specialised Horizon Code-Switch model is designed for conversations that naturally mix languages.
Shunya's intelligence layer includes custom redaction and profanity masking for supported batch workflows.
Yes. Video files can be submitted through the batch transcription workflow.
The transcript itself can be indexed and searched by your application. Word and segment timestamps can then connect search matches back to positions in the original video.
Get started
Don't judge video transcription from our prepared clip. Upload something real. Try an interview, webinar,podcast, meeting recording, lecture, customer story, multilingual video, or noisy real-world recording.See the transcript, speakers, timestamps, and structured output for yourself.