Examples
- Hi, I wanted to check whether my payment was processed yesterday.
- Can you move my appointment to Friday afternoon?
- Mera order अभी तक deliver नहीं हुआ.
Horizon · WebSocket streaming
Convert live speech into text while the conversation is still happening for voice agents, customer support, agent assist, live captions, meetings, call routing, search, and real-time AI applications.
Stream audio to Shunya Horizon over WebSocket and receive partial and final transcripts with multilingual automated speech recognition, telephony audio support, speaker-aware workflows, and structured output for downstream applications.
I wanted to check whether my payment was processed yesterday.
I wanted to check whether my payment was processed yesterday.
Partial transcript, then final transcriptReal-time transcription
Traditional transcription starts after a recording is complete. Real-time transcription works differently. The application sends audiocontinuously while the person is speaking. Horizon processes that live audio and returns text as the conversation progresses.
This makes live speech usable by software before the conversation ends.
| Capability | Horizon |
|---|---|
| Real-Time Transcription | |
| WebSocket Streaming | |
| Partial Transcripts | |
| Final Transcripts | |
| Live Microphone Audio | |
| PCM Streaming | |
| float32 Streaming | |
| μ-law | |
| A-law | |
| Telephony Audio | |
| Multilingual Recognition | |
| Automatic Language Detection | |
| Voice-Agent Integration | |
| Agent Assist | |
| Live Captions | |
| Call Routing | |
| Real-Time Search |
Try it yourself
Do not test real-time transcription only with a carefully prepared sentence. Try speech that resembles your production environment.
Examples
Test with
Partial transcripts
A speaker begins: “I wanted to check...” Horizon can return an interim transcript while more audio is still arriving.
Partial transcripts allow applications to begin working with the conversation before the full utterance is final. Shunya explicitly documents interim/partial transcription for live streaming.
Final transcripts
Partial text can change as additional audio provides more context. Final transcripts represent completed segments.
Partial
Applications can treat final transcripts differently from partial text.
Final
This distinction matters when building real-time systems.
Sub-500 ms streaming performance
Shunya publishes sub-500 ms streaming transcription and first-token performance forreal-time Horizon workflows—useful when transcription must appear during speech rather thanafter a recording finishes.
Sub-500 msStreaming transcription / first-token performance for real-time Horizon workflows.
Use cases
ImportantActual end-to-end application latency also depends on network, audio chunking, client implementation, model configuration, and downstream application logic. The published figure should not be interpreted as a guarantee for every complete application stack.
Latency also depends on
WebSocket streaming
Recorded transcription typically uses an HTTP request. Real-time transcription uses a persistent WebSocket workflow.
The WebSocket remains part of the ongoing voice interaction rather than representing one completed file upload.
Live microphone audio
Real-time audio may come from a browser microphone, mobile microphone, desktop application, voice assistant, call application, or meeting application.
Use for
Speech can become an input method in the same way typing is.
Live audio formats
Shunya documents live audio representations including int16 PCM, float32, μ-law, and A-law.
For general recording and real-time audio.
For pre-processed audio streams.
For telephony and SIP workflows.
For G.711 telephony workflows.
This lets the same streaming platform work across different real-time environments.
Real-time phone call transcription
Contact-center applications often need transcription while the customeris still speaking. Shunya explicitly documents this workflow.
The application can use the transcript before the customer hangs up.
μ-law and A-law
Shunya documents native streaming support for G.711 μ-law, common in telephony environments, and G.711 A-law, common in SIPand telephone infrastructure. Shunya documents an 8 kHz telephony workflow for these codecs.
G.711 μ-law
Common in telephony environments.
G.711 A-law
Common in SIP and telephone infrastructure.
8 kHzFor supported μ-law and A-law frames, Shunya does not require a separate resampling or codec-conversion layer before transmission.
Voice agents
An AI voice agent first needs to understand what the user says. Real-time automated speech recognition is typically the listening layer in a voice-agent pipeline.
Speech, then text, then intelligence, then speech. Real-time transcription enables the listening side of that loop.
Agent assist
A human support agent can benefit from seeing what the customer says in real time. Shunya explicitly lists agent assistamong real-time call-transcription use cases. The surrounding application controls what assistance is displayed.
Possible downstream uses
Live captions
Real-time transcription can power live caption experiences. The application controls caption layout and presentation. Horizon provides the live text.
Possible use cases
Call routing
Shunya explicitly lists call routing among live transcription use cases. The routing rules remain in the surrounding application.
Traditional IVR
Conversational system
Knowledge retrieval
Shunya's real-time call material explicitly includes knowledge retrieval and real-time search among streaming use cases.
The transcript can act as input to
Real-time search
A live transcript can be indexed or queried as speech arrives. The search engine itself remains external. Horizon supplies the live text input.
Potential uses
Real-time meetings
Streaming transcription can support live meeting experiences. For complete post-meeting intelligence, the finished recording can also be processed through batch transcription.
Potential downstream features
Multilingual real-time transcription
Horizon supports multilingual speech recognition. A real-time application does not need a completely separate architecture for each supported market.
Examples include
Automatic language detection
Horizon supports automatic language detection. When language is known, anapplication can explicitly configure the expected language.
For short clips, locking the known language can reduce ambiguity because there is less speech context available for detection.
Code-switched live speech
Horizon includes specialised Code-switch workflows for supported mixed-language conversations.The relevant language pair should still be verified against the deployed configuration.
Examples
Do not assume every possible language combination is supported by a dedicated Code-switch model.
Speaker diarization
A conversation is more useful when each participant remains attached to their words. Shunya documents speaker diarization as part of its broader transcription stack.
Without diarization
I haven't received my refund yet. I'll check that for you.
With diarization
SPEAKER_00 I haven't received my refund yet.
SPEAKER_01 I'll check that for you.
Use for
Speaker identification
Diarization tells you that different people spoke. Speaker identification can match registered voice profiles. Plan for approximately 5–15 seconds of clean solo speech when registering a speaker profile.
Before
I'll review that for you.
After
I'll review that for you.
Unknown participants can remain anonymous labels.
Word-level timestamps
Verbose transcript output can include word start time, word end time, and a confidence score. Timing makes transcripts operational rather than merely readable.
Use for
Real-time vs batch
Streaming and batch transcription serve different jobs. Richer conversation intelligenceis often available on completed recordings—not automatically live during streaming.
Streaming
Batch
Post-call intelligence
For supported batch workflows, Shunya documents speaker structure, summaries, intent, sentiment, emotion, keyterms, redaction, and word timestamps.
Understand now. Analyse more deeply afterwards.
Customer support
Possible real-time downstream use includes agent assist, routing, search, knowledge retrieval, and avoice agent. After the call, the recording can be used for summaries and analytics.
Possible real-time downstream use
Contact centers
Shunya positions live call transcription for contact-center environments. One transcript can support several applications around the same conversation.
Possible downstream workflows
Dictation and voice input
Streaming transcription can also support direct voice input. The transcript can be edited, stored, searched, or processed downstream.
Possible experiences
Audio chunking and application design
An application does not upload an entire finished recording. It sends audio progressively. Application designtherefore matters. End-to-end responsiveness depends not only on the speech model but alsoon how audio is captured, transmitted, and consumed.
Developers
Live audio formats Shunya documents
Input
Configure according to workflow
Send the transcript into
Choose the right Horizon model
One real-time API architecture can route to the recognition path that best matches the conversation.
Horizon Universal
For supported global languages. Examples include English, Spanish, French, Japanese, and Korean.
Horizon Indic
For Indian-language and regional speech.
Horizon Code-Switch
For supported mixed-language conversations.
Horizon Medical
For specialised medical speech.
Real-world real-time use cases
Live speech wherever text is needed before the audio ends.
Understand users while they speak.
Provide human agents with live transcript context.
Display spoken words as text.
Use speech as input to support workflows.
Route calls according to natural spoken requests.
Turn the caller's speech into a query.
Search systems while the conversation continues.
Create live meeting transcripts.
Let users speak queries.
Convert speech directly into editable text.
Send live human speech into an LLM or agent.
Built for production real-time transcription
Production streaming transcription may need persistent connections, partial and final transcripts, telephonyaudio, multilingual recognition, speaker-aware workflows, and batch enrichment after the call.
| Requirement | Support |
|---|---|
| Real-time transcription | |
| WebSocket | |
| Partial Transcripts | |
| Final Transcripts | |
| Sub-500 ms Published Streaming Performance | |
| int16 PCM | |
| float32 | |
| μ-law | |
| A-law | |
| SIP / PSTN Workflow | |
| 8 kHz Telephony Workflow | |
| Multilingual Recognition | |
| Automatic Language Detection | |
| Code-switch Platform | |
| Speaker Diarization | |
| Speaker Identification | |
| Word Timestamps | |
| Voice Agents | |
| Agent Assist | |
| Live Captions | |
| Batch Post-call Intelligence |
Send live text into
Questions
Real-time transcription converts live speech into text while the person is still speaking instead of waiting for a complete recording.
Yes. Horizon supports WebSocket-based streaming automated speech recognition.
Yes. Partial or interim transcripts can be returned while speech continues.
Yes. Transcript segments are finalised as the speech progresses.
Shunya publishes sub-500 ms streaming and first-token performance for current Horizon streaming workflows.
No. End-to-end latency also depends on networking, client audio handling, downstream processing, and application architecture.
Shunya uses a WebSocket streaming workflow for live speech.
Shunya documents int16 PCM, float32, μ-law, and A-law for live streaming.
Yes. Shunya explicitly supports live call transcription.
Yes. G.711 μ-law is documented for native telephony streaming.
Yes. G.711 A-law is also documented.
Shunya documents an 8 kHz telephony workflow for these codecs.
For supported μ-law and A-law streaming frames, Shunya does not require a separate resampling or codec-conversion step before transmission.
Yes. Voice agents are explicitly listed among streaming use cases.
Yes. Agent assist is explicitly listed.
Yes. Live captions are explicitly listed.
Yes. Call routing is explicitly listed as a real-time transcription use case.
Yes. The live transcript can be passed into a knowledge or search application.
Horizon supports multilingual recognition across supported languages.
Yes. Automatic language detection is supported when the language is not explicitly fixed.
Horizon Code-Switch supports specialised workflows for supported mixed-language combinations.
Not every language pair is supported. Verify the combination you need against your deployed configuration.
Speaker diarization is part of Shunya's transcription stack and can create speaker-labelled transcript segments.
Registered speaker profiles can be used for speaker identification where configured.
Yes. Verbose transcript output can include word-level timing and confidence.
Richer intelligence such as intent, sentiment, emotion, summary, and redaction is primarily available through batch and post-call workflows. Do not assume every enrichment is available live during streaming.
Yes. The completed recording can be processed through the batch workflow for richer conversation intelligence.
Streaming processes live audio over WebSocket and returns partial and final transcripts. Batch processes completed recordings over HTTP and can include richer post-call intelligence.
Get started
Don't wait until the conversation is over. Stream microphone, application, or telephony audio into Horizon. Receive partialtranscripts while the speaker continues. Receive final transcripts as segments complete. Send that live text into a voiceagent, agent-assist tool, caption interface, search system, call router, or AI application.