Try
Real call scenarios
- A customer-support call
- Sales conversation
- Payment inquiry
- Service request
- Complaint
- Appointment call
- Recorded IVR interaction
- Multilingual conversation
- Noisy mobile call
Horizon · Call Transcription
Turn customer calls into accurate, speaker-aware transcripts for contact centers, agent assist, summaries, quality monitoring, analytics, search, compliance workflows, and conversational AI. Use Shunya Horizon to transcribe live or recorded calls across 204+ languages - including native telephony audio, speaker labels, word-level timestamps, multilingual conversations, and post-call intelligence.
Hi, I made the payment yesterday but I still haven't received the confirmation.
Speaker diarization · HorizonCall transcription
Calls are difficult speech-recognition environments. Customers speak from noisy places. Agents use different headsets. Telephone networks compress the audio. People interrupt each other. Names and account numbers appear without warning. And customers frequently mix languages inside the same conversation. Horizon is designed for the audio businesses actually receive.
204+
Languages across the current Horizon platform
3.10%
Composite WER across 8 OpenASR benchmarks
Sub-500ms
Streaming transcription on the current product platform
240+
Concurrent streams per GPU shown on the current Horizon product page
| Capability | Horizon · Call Transcription |
|---|---|
| Live Calls | |
| Recorded Calls | |
| Phone-line Audio | |
| SIP / PSTN Audio | |
| μ-law Input | |
| A-law Input | |
| Speaker Diarization | |
| Speaker Identification | |
| Word-level Timestamps | |
| Multilingual Calls | |
| Code-switching | |
| Intent & Sentiment | |
| Call Summaries | |
| Redaction |
Try it yourself
Don't evaluate call transcription using studio audio. Try a real customer-support call,
sales conversation, payment inquiry, complaint, or multilingual recording.
Try
Example
CUSTOMER Hi, I made the payment yesterday but I still haven't received the confirmation.
AGENT I'll check that for you. Could you share your transaction number?
CUSTOMER Yes, it's 48217.
Try workflow
Built for phone audio
Telephone audio creates its own ASR challenges. Crosstalk is common - agents and customers may interrupt each other.
Traditional telephony may operate at only 8 kHz.
SIP and PSTN systems commonly use G.711 codecs.
The customer's phone and the agent's headset can have very different quality.
Packet loss, distortion, jitter, and compression can affect the audio.
Background noise
Shunya positioning
Real call audio
The goal is not speech recognition that works only when the recording is clean.
Native SIP / PSTN audio
Shunya's streaming ASR directly supports the two G.711 codecs widely used by telephone networks.
μ-law
Common across North American and many Indian PSTN systems.
A-law
Common in European telephony and some SIP environments.
8 kHz
Native telephony workflow sample rate
Phone call
No separate resampling or codec-conversion layer is required before sending supported μ-law or A-law frames to the streaming endpoint.
Real-time call transcription
Contact-center applications often need the transcript before the call ends. Horizon supports streaming speech recognition over
WebSocket. Text arrives while speech continues, and segments are finalised as the conversation progresses.
Sub-500ms
Current published first-token streaming performance
Live call
Use live call transcription for
The application does not have to wait until the customer hangs up before it starts using the transcript.
Recorded call transcription
Not every workflow needs real-time transcription. Recorded calls can be processed through Horizon's batch API.
Call recording
Optional intelligence
Best for
Live
After
Who said what
A call transcript becomes much more useful when the two sides remain distinct.
Without diarization
I haven't received my refund yet. I'll check that for you. Can you share your transaction number?
With diarization
SPEAKER_00 I haven't received my refund yet.
SPEAKER_01 I'll check that for you. Can you share your transaction number?
2–6
Distinct voices where diarization generally performs strongest - a standard agent-customer call fits naturally
Speaker diarization
Shunya supports diarization as working with any number of speakers.
Identify the agent
Diarization tells you that two people spoke. Speaker identification can resolve registered voices into names.
Before
I'll review the transaction for you.
After
I'll review the transaction for you.
01
Use a clean reference clip.
02
Keep profiles inside the relevant project.
03
Separate agent and customer turns.
04
Resolve matching speaker to a name.
5–15 seconds
Shunya currently recommends approximately this much clean solo speech for speaker registration
Contact-center use: agents can be registered; customers can remain anonymous - producing transcripts labelled Customer and Priya rather than SPEAKER_00 and SPEAKER_01.
Call accuracy
A customer call is very different from a studio recording. Shunya's broader benchmark suite includes multiple difficult
speech environments - including Earnings22, which contains challenging earnings-call speech.
5.83%
Horizon WER on Earnings22
| Benchmark | Domain | Horizon WER |
|---|---|---|
| LibriSpeech Clean | Clean read speech | 0.71% |
| TED-LIUM | Presentations | 1.43% |
| AMI | Multi-party meetings | 4.19% |
| GigaSpeech | Podcasts / YouTube | 4.99% |
| Earnings22 | Earnings calls | 5.83% |
Read this fairly
It provides evidence that Horizon is evaluated on difficult conversational and call-style audio. It does not guarantee 5.83% WER on every customer phone call. Call accuracy changes with:
Your own calls remain the final test.
Multilingual calls
A call may begin in English, switch to Hindi, use an English product name, then return to Hindi. The model should follow the way the customer actually talks.
मैंने कल payment किया था but confirmation अभी तक नहीं आया.
मेरा order deliver हो गया but app में status pending दिखा रहा है.
Horizon Universal
For broad global calls.
Horizon Indic
For Indian-language conversations.
Horizon Code-Switch
For naturally mixed speech.
Use for
Horizon Code-Switch
Call intent
A transcript tells you the words. Intent classification can help identify the reason behind the call.
Example intents
Call
Or a configured business-specific category.
Constrain the output
complaintinquiryservice_requestcomplimentUse intent for
Call sentiment & emotion
Two customers may say similar things with very different emotional signals.
Sentiment
With an associated score and explanation.
Emotion
Example
Use for
Important. These are model-generated classifications. They should support human and business workflows rather than be treated as infallible descriptions of a person's internal state.
Call summaries
The full transcript preserves detail. A summary gives the next person a faster view of the interaction.
Call
Use call summaries for
Example workflow
Then
Word-level timestamps
Verbose transcription can return per-word start time, end time, and confidence score.
Search
Call recordings become searchable instead of requiring manual listening from the beginning.
Keyterm normalisation
Contact-center calls contain organisation-specific vocabulary. A transcript can be generally readable and still become operationally useless if the one incorrect word is the customer's product, plan, or service name.
Contact-center vocabulary
Example
Configure
Redaction & profanity
Calls may contain information that should not appear unprotected across downstream systems.
Shunya supports configured keyword masking and profanity hashing.
Examples
Original
Please give me your card number and OTP.
Redacted
Please give me your **** and ****.
Profanity hashing
****
Important. Keyword redaction is one layer. Organisations should still apply appropriate access controls, retention policies, data governance, recording consent, and compliance processes according to their call environment.
Live now, enrich after
Shunya currently documents its intelligence layer as working with batch ASR. A technically accurate
contact-center architecture separates real-time transcription from post-call enrichment.
During the call
Understand what the customer is saying now.
After the call
Run the recording through the batch workflow, then enable intelligence features.
This prevents the architecture from implying that every intelligence feature runs live when Shunya currently documents them as batch enrichments.
Call analytics
Once calls become structured text, they can be aggregated and analysed. Call transcription turns voice conversations into data organisations can measure.
Calls
Analytics
Connect to
From audio to queryable data
Contact centers
Shunya explicitly positions Horizon for contact-center environments.
Turn live customer speech into text.
Use the transcript to retrieve knowledge or help agents during calls.
Process recorded calls for review and analysis.
Understand common call reasons.
Track broad conversation sentiment.
Create shorter records of completed calls.
Use one speech platform across supported languages.
Process telephony codecs directly through streaming ASR.
Contact center flow
Banking & financial calls
Financial calls often contain precise terminology and identifiers. Transcription and redaction features support the technical workflow - organisations remain responsible for legal, regulatory, consent, retention, and security requirements in their environment.
Financial calls often contain
Use transcription for
Redaction. Organisation-defined sensitive terms can be masked before transcripts enter downstream systems.
Call transcription API
Use batch transcription for recorded calls, post-call analytics, summaries, and intelligence. Use WebSocket streaming for live calls, voice agents, and agent assist. Choose the model according to language - for example horizon-indic for supported Indian-language calls.
curl -X POST https://asrv2prod.shunyalabs.ai/v1/audio/transcriptions \
-H "Authorization: Bearer $ACCESS_TOKEN" \
-F "file=@call.wav" \
-F "model=horizon-indic" \
-F "enable_diarization=true" \
-F "enable_intent_detection=true" \
-F "enable_sentiment_analysis=true" \
-F "enable_summarization=true" \
-F "response_format=verbose_json"Precise search
Choose the right call model
Global calls
Use for broad international-language coverage.
Best for:
Indian-language calls
Use for supported Indian languages and regional speech.
Best for:
Mixed-language calls
Use when conversations naturally switch languages.
Best for:
Healthcare-oriented calls
Use the appropriate domain-specific speech model where the use case and language support fit.
Best for:
Model follows the call
Built for production
Recording a call is easy. Turning every conversation into reliable, searchable, structured information is harder. Productioncall transcription needs telephony codecs, streaming, speaker structure, intelligence, APIs, multilingualsupport, redaction, and infrastructure that can handle call volume.
| Requirement | Horizon · Call Transcription |
|---|---|
| Live Calls | |
| Recorded Calls | |
| SIP / PSTN Workflow | |
| μ-law Input | |
| A-law Input | |
| 8 kHz Telephony | |
| 204+ Languages | |
| Speaker Diarization | |
| Speaker Identification | |
| Word-level Timestamps | |
| Multilingual Calls | |
| Code-switching | |
| Intent Detection | |
| Sentiment Analysis | |
| Emotion Analysis | |
| Summarisation | |
| Keyterm Normalisation | |
| Redaction | |
| Streaming API | |
| Batch API | |
| Enterprise Deployment |
Production flow
Questions
Call transcription uses automatic speech recognition to convert telephone conversations into written text.
Yes. Horizon is explicitly positioned for contact-center and phone-line audio.
Yes. Horizon provides WebSocket streaming for live speech recognition.
Yes. Recorded calls can be processed through the batch transcription API.
Yes. Shunya supports native streaming support for the G.711 codecs commonly used in telephone networks.
Yes. The streaming endpoint accepts G.711 μ-law audio natively.
Yes. The streaming endpoint also accepts G.711 A-law audio natively.
Shunya's SIP/PSTN documentation specifies 8 kHz for the supported μ-law and A-law workflows.
For supported μ-law and A-law streaming frames, Shunya states that no separate resampling or codec conversion is required before transmission.
Yes. Speaker diarization separates different voices into speaker-labelled segments.
Yes. Register the agent's voice profile and combine speaker identification with diarization.
Shunya currently recommends approximately 5–15 seconds of clean solo speech.
Yes. Verbose JSON output can return word-level start time, end time, and confidence scores.
Shunya's batch intelligence layer includes intent detection and allows the application to constrain output to configured intent choices.
Yes. Batch sentiment analysis returns a positive, negative, or neutral label, a numeric score, and an explanation.
Emotion diarization can add dominant emotion to transcript segments.
Yes. Summarisation is available through the batch intelligence layer.
Yes. Custom keyword redaction can mask configured terms, and profanity hashing can mask profane language.
Yes. Horizon supports broad multilingual recognition, while specialised models are available for Indian and code-switched conversations.
Shunya publishes 5.83% WER on Earnings22, an earnings-call benchmark. It is not a universal contact-center WER: customer-service phone calls may have different codecs, noise, accents, and speaking conditions.
Shunya's current Horizon product page publishes 240+ concurrent streams per GPU as a platform metric. Actual deployment capacity depends on the model, infrastructure, audio characteristics, and deployment configuration.
Yes. Shunya provides HTTP batch transcription and WebSocket streaming, including native telephony-audio workflows.
Get started
Don't judge call transcription from our prepared recording. Test one of yours. Try a customer-support call, complaint, noisymobile conversation, Hinglish call, payment inquiry, sales conversation, service request, or real PSTN recording.See the transcript, speakers, timestamps, intent, sentiment, and post-call intelligence for yourself.