01
Add your audio
Upload a recorded file, provide an audio source, or send live audio through a streaming connection.
Horizon · Audio to Text
Convert recorded or live audio into accurate, structured text for meetings, calls, interviews, podcasts, media, analytics, and AI applications. Upload an audio file or send an audio stream to Shunya Horizon and turn speech into searchable transcripts, speaker-labelled conversations, timestamps, and structured intelligence.
Hi, I'm calling about an order I placed yesterday. Sure - could you share the order number with me? Yes, it's 48217.
Speaker diarization · HorizonAudio to Text
Audio contains more than words. Real recordings include different speakers, accents, background noise, interruptions, pauses, technical terminology, and sometimes more than one language. A production audio-to-text system needs to preserve enough structure to make the transcript useful after the audio has been processed.
204+
Languages supported across the Horizon platform
3.10%
Composite WER across 8 OpenASR benchmarks for Horizon Universal
Sub-500ms
Streaming transcription latency for real-time speech
#1
OpenASR composite leaderboard performance
| Capability | Horizon · Audio to Text |
|---|---|
| Recorded Audio | |
| Live Audio Streaming | |
| Automatic Language Detection | |
| Multiple Speakers | |
| Speaker Diarization | |
| Word-level Timestamps | |
| Confidence Scores | |
| Punctuation | |
| Multilingual Transcription | |
| Conversation Intelligence | |
| Self-hosted Deployment |
Try it yourself
A prepared benchmark clip cannot tell you how an ASR system will perform on yourcustomer calls, meetings, accents, microphones, or vocabulary.
Audio → transcript
Audio-to-text starts with speech recognition, but useful transcription should preserve the structure of the recording wherever possible.
01
Upload a recorded file, provide an audio source, or send live audio through a streaming connection.
02
Horizon recognises the spoken language and converts speech into text.
03
Add speakers, timestamps, confidence information, and optional conversation intelligence.
04
Search it. Analyse it. Summarise it. Caption it. Store it. Send it into another AI workflow.
One transcription path
Input: audio. Processing: speech recognition plus transcript structure. Output: text, speakers, timestamps, language, and optional intent, sentiment, and summary fields.
Built for real audio
Production recordings come from the environments where people actually speak.
Calls may come from traffic, offices, public spaces, homes, or low-quality phone connections.
Audio quality varies between mobile phones, laptops, headsets, conference systems, and telephony infrastructure.
Meetings, interviews, calls, and podcasts often contain speaker changes, interruptions, and overlapping speech.
The same language can sound very different depending on region, community, and speaker.
Product names, technical terminology, financial terms, medical vocabulary, names, and abbreviations can make transcription harder.
People hesitate, restart sentences, shorten words, speak over each other, and rarely sound like scripted benchmark recordings.
Audio to Text needs to work on the audio your application receives - not only on clean samples.
Who said what
A plain transcript can tell you what was said. Speaker diarization tells you who spoke when.
Example
SPEAKER_00 Hi, I'm calling about an order I placed yesterday.
SPEAKER_01 Sure. Could you share the order number with me?
SPEAKER_00 Yes, it's 48217.
Instead of returning the entire conversation as one block, diarization assigns speaker labels to individual transcript segments.
Useful for
Speaker identification
Where configured, speaker profiles can also be used to resolve anonymous speaker segments against known voices.
Audio → transcript
Word-level timing
For many applications, the words alone are not enough. You also need to know where those wordsappear in the original recording. Horizon can return word-level timing information.
Beyond plain text
Audio often enters a business workflow because someone eventually needs to understand what happened in the conversation. Thetranscript is the first layer. Horizon can add optional intelligence fields to the same batch transcription workflow.
Identify who spoke when.
Match transcript segments to configured speaker profiles.
Identify what the speaker is trying to accomplish.
Analyse whether the conversation or segment is positive, negative, or neutral.
Detect the dominant emotion associated with speech segments.
Reduce a long recording into a shorter structured overview.
Normalise important words and domain-specific terminology.
Generate translated transcript output when required.
Mask profane language in the transcript.
Hide specified sensitive terms or phrases from transcript output.
Return timing information for individual words.
Structured output
The output can become structured conversation data instead of remaining a raw block of text.
Multilingual audio
Audio processing becomes more complicated when every language requires a separate transcriptionprovider or separate model pipeline. Horizon is built around multilingual speech recognition.
204+
Languages supported across the Horizon platform
Automatic language detection
When the language is not known in advance, the transcription workflow can detect the spoken language automatically.
Language-specific models
For applications that know the language or require specialised performance, teams can choose an appropriate Horizon model.
Code-switched conversations
Example: I made the payment yesterday लेकिन अभी तक confirmation नहीं आया.
Rather than forcing every recording into a strictly single-language workflow, Horizon includes specialised support for multilingual and code-switched speech.
One audio platform
Recorded + live audio
Audio-to-text workflows generally fall into two categories: batch transcription for existing recordings and streaming transcription for live audio.
Batch
Send a complete recording and receive the transcription after processing.
Best suited for:
Streaming
Send audio continuously through a persistent connection and receive interim and final transcription as people speak.
Best suited for:
| Requirement | Batch | Streaming |
|---|---|---|
| Existing audio files | ||
| Live microphone audio | ||
| Complete transcript | ||
| Partial transcript | ||
| Post-call analytics | ||
| Voice agents | ||
| HTTP API | ||
| WebSocket |
Audio formats
Audio may come from a mobile recording, meeting platform, podcast, phone system, browser, or another application.Horizon supports common formats for recorded audio. Streaming workflows can use audiorepresentations designed for real-time applications and telephony.
| Batch audio | Description |
|---|---|
| WAV | Uncompressed audio commonly used for high-quality recordings. |
| MP3 | Widely used compressed audio for general recordings and media. |
| M4A | Common in mobile and recorded voice workflows. |
| OGG | Compressed open audio format. |
| FLAC | Lossless compressed audio. |
| WebM | Common in browser-based and web-recording workflows. |
| Live audio | Description |
|---|---|
| PCM | General real-time audio. |
| μ-law | Telephone and SIP audio. |
| A-law | Telephony infrastructure. |
One transcription workflow
Real-world use cases
Convert discussions into searchable transcripts with speaker labels and timestamps.
Turn recorded conversations into structured text for quality monitoring, summaries, analytics, and search.
Create searchable transcripts, captions, show notes, and downstream content from recorded episodes.
Convert research, journalism, hiring, and customer interviews into structured written material.
Generate transcripts and subtitles for broadcasts, recorded videos, interviews, and media archives.
Turn lectures, lessons, seminars, and classroom recordings into searchable text.
Convert spoken documentation and recorded conversations into structured transcription workflows where appropriate.
Process customer conversations for searchable records, analysis, summaries, and defined redaction requirements.
Turn mobile recordings and spoken notes into written information that can be searched or processed.
Developers
Horizon provides an audio transcription endpoint for programmatic processing. Send recorded audio to POST /v1/audio/transcriptions or use streaming connections for live audio.
{
"language_code": "auto",
"file": "meeting.wav",
"enable_diarization": true
}Input
Or provide an audio source for batch processing.
Model
Select the appropriate Horizon model for the language and application.
Configure
Output
Basic workflow
Choose the right STT
No single model or configuration is ideal for every language and application. The Horizon familyincludes specialised speech-recognition options for different requirements.
Model
Multilingual speech recognition for broad language coverage and everyday conversational audio.
Model
Purpose-built recognition for Indian languages and speech patterns.
Model
Designed for conversations that naturally move between languages.
Model
Specialised speech recognition can be used where terminology, deployment, or industry requirements demand a more focused model.
One platform, different audio: Choose the model according to the audio instead of forcing every recording through the same recognition path.
Built for production
Converting one MP3 file to text is easy to demonstrate. Production audio workflows need much more. They needmultilingual recognition, speaker separation, timestamps, structured data, APIs, high-volumeprocessing, real-time support, analytics, and deployment control.
| Requirement | Horizon |
|---|---|
| Audio File Transcription | |
| Real-time Audio | |
| Multiple Audio Formats | |
| Automatic Language Detection | |
| Multilingual STT | |
| Speaker Diarization | |
| Speaker Identification | |
| Word-level Timestamps | |
| Intent Detection | |
| Sentiment Analysis | |
| Emotion Detection | |
| Summarisation | |
| Redaction | |
| Code-switching | |
| API Integration | |
| Self-hosted Deployment |
Audio → AI
The transcript becomes machine-readable information that another system can actually use.
Questions
Audio to Text is speech-recognition technology that converts spoken audio from recordings or live streams into written text.
Yes. Horizon supports batch transcription for recorded audio files and streaming transcription for live audio.
The current batch transcription workflow supports formats including WAV, MP3, M4A, OGG, FLAC, and WebM.
Yes. Speaker diarization can assign speaker labels to transcript segments so multi-speaker conversations are easier to follow.
Yes. Horizon can return segment timing and word-level timestamps.
Automatic language detection can be used when the language is not specified in advance.
Yes. Horizon provides multilingual speech recognition, while the specialised Horizon Code-Switch model is designed for conversations that move naturally between languages.
Yes. Horizon supports real-time transcription through streaming WebSocket connections.
The optional Horizon intelligence layer supports summarisation for batch transcription workflows.
Yes. Optional intelligence features include sentiment and intent analysis alongside the transcript.
Custom redaction and profanity-masking features can be enabled for supported transcription workflows.
Yes. Recorded audio can be sent to the Horizon transcription API, while live audio can be handled through streaming connections.
Shunya supports self-hosted deployment options for organisations that need greater control over infrastructure and audio data.
Get started
Don't test transcription with our recording. Test yours. Upload a meeting, customer call, interview,podcast, voice note, multilingual conversation, or noisy real-world recording. See thetranscript, speakers, timestamps, and structured output for yourself.