Upload
Use a real meeting
- A team discussion
- Sales meeting
- Customer call
- Interview
- Research session
- Planning meeting
- Remote conference
- Recorded workshop
Horizon · Meeting Transcription
Turn meetings into accurate, speaker-aware transcripts for notes, search, summaries, review, collaboration, analytics, and AI workflows. Use Shunya Horizon to transcribe live or recorded meetings across 204+ languages, separate speakers, preserve word-level timing, and enrich completed meeting transcripts with summaries and conversation intelligence.
Let's move the launch to Thursday so the mobile team has another day for testing.
Speaker diarization · HorizonMeeting transcription
Meetings are difficult speech-recognition environments. Several people speak. Someone interrupts. Participants use different microphones. One person joins remotely. Another speaks from across the room. Product names, numbers, acronyms, and technical terminology appear without warning. And multilingual teams may change languages during the same discussion. Horizon is built to turn that conversation into structured text.
204+
Languages supported across Horizon
3.10%
Composite WER across 8 OpenASR benchmarks
4.19%
Published WER on the AMI meeting benchmark
Sub-500ms
Streaming first-token latency
| Capability | Horizon · Meeting Transcription |
|---|---|
| Recorded Meetings | |
| Live Meetings | |
| Multiple Speakers | |
| Speaker Diarization | |
| Speaker Identification | |
| Word-level Timestamps | |
| Confidence Scores | |
| Multilingual Meetings | |
| Code-switching | |
| Meeting Summaries | |
| Intent & Sentiment | |
| Keyterm Normalisation | |
| Private Deployment |
Try it yourself
The best way to evaluate meeting transcription is with an actual meeting.
Upload
Example meeting
SPEAKER_00 Let's move the launch to Thursday so the mobile team has another day for testing.
SPEAKER_01 That works for engineering, but we need final copy by Wednesday morning.
SPEAKER_02 I'll confirm the content by tomorrow afternoon.
Model
Transcribe the meeting with the Horizon model that matches the conversation.
Enable
After the meeting
Meeting accuracy
Meeting audio is different from clean read speech. The AMI Meeting Corpus includes recorded meetingconversations and is one of the benchmarks included in Shunya's published OpenASR evaluation.
4.19%
Horizon WER on AMI
54%
Relative improvement over the best published baseline shown in Shunya's benchmark comparison
| Model | AMI WER |
|---|---|
| Horizon | 4.19% |
| Best published baseline in comparison | 9.12% |
| NVIDIA Canary-Qwen | 9.65% |
Why meeting audio is hard
Read this fairly
It does not guarantee 4.19% WER on every meeting. Actual accuracy depends on:
The benchmark provides evidence. Your own meetings provide the final test.
Who said what
Consider this transcript: We should launch Thursday. I agree, but design needs another day. That's fine, I'll update the timeline. The wordsare readable. But who agreed? Who needs another day? Who owns the timeline? Speaker diarization adds that structure.
Speaker diarization
SPEAKER_00 We should launch Thursday.
SPEAKER_01 I agree, but design needs another day.
SPEAKER_02 That's fine. I'll update the timeline.
Speaker diarization
Shunya's diarization works with meetings containing multiple speakers and is designed to preserve speaker identity at the transcript-segment level.
Speaker identification
Diarization answers who spoke when. Speaker identification can go one step further. When known participantshave registered voice profiles, anonymous labels can be resolved into speaker names.
Without speaker identification
The campaign goes live next Monday.
With speaker identification
The campaign goes live next Monday.
01
Use a clean reference clip.
02
Build a team or project voice library.
03
Separate the meeting by speaker.
04
Resolve matching speakers by name.
5-15 seconds
Recommended length of clean solo speech for registering a speaker
Unrecognised participants can remain anonymous SPEAKER_XX labels. This makes speaker identification useful without requiring every possible meeting participant to be pre-registered.
Multiple speakers
Speech-recognition models can perform well on individual speakers and still struggle once several peoplebegin interacting. Shunya's meeting workflow can preserve speaker-level segmentation.
2-6
Diarization works with any number of speakers and typically gives its best results with this many distinct voices
Designed for
What makes it harder
Clearer audio → clearer speaker separation. Meeting transcription quality depends on both ASR accuracy and the quality of speaker separation.
Word-level timestamps
A meeting transcript becomes much more useful when text can be connected back to the originalrecording. Horizon can return per-word start time, end time, and confidence score.
Search
The transcript becomes an index into the meeting rather than a separate document disconnected from the audio.
Live meeting transcription
Some meeting workflows only need the transcript afterward. Others need text in real time. Horizon supports streaming transcription for live speech applications. Text appears while the speaker continues, and segments are finalised as the conversation progresses.
Sub-500ms
Published first-token streaming latency
Live audio
Use live transcription for
Live now, enrich after
Not every meeting feature needs to run while someone is speaking. Shunya's intelligencelayer runs on batch ASR. That creates a useful two-stage meeting workflow.
During the meeting
Streaming transcription keeps text moving while the conversation is still happening.
After the meeting
Run the completed audio through batch transcription, then enable the intelligence layer.
This separation keeps live transcription and post-meeting enrichment distinct. The enrichments run as batch capabilities on the completed recording.
Meeting summaries
A transcript preserves detail. A summary helps people understand the conversation quickly. Shunya's intelligence layer includes summarisation for completed transcription workflows.
Meeting audio
Use summaries for
Important distinction
If your application also needs the structures below, the transcript and summary can be passed into an LLM or workflow designed to extract them.
From speech to next step
From meeting to intelligence
Shunya's batch intelligence features can enrich a completed meeting transcript. A meeting becomes structured information that other systems can use.
Who spoke when?
Which registered participant was speaking?
What dominant emotion appears in each segment?
What type of request or objective appears in the conversation?
Is the conversation positive, neutral, or negative?
What happened across the meeting?
Standardise important domain terminology.
Where exactly was each word spoken?
Mask organisation-defined sensitive terms.
Replace profane words in transcript output.
Meeting audio
Multilingual meetings
Global and Indian teams often switch languages naturally. A participant may present in English. Another may ask a question in Hindi. Technical terminology may remain English even when the surrounding conversation changes language. Shunya Horizon is built for multilingual and code-switched speech.
The API migration is scheduled for Friday.
ठीक है, लेकिन deployment से पहले testing complete हो जाएगी?
Yes, QA should finish by Thursday evening.
204+
Global language coverage
Specialised model option
Global multilingual meetings.
Specialised model option
Meetings primarily in Indian languages.
Specialised model option
Meetings that naturally mix languages.
One meeting
Preserve the conversation without forcing every speaker to use a single language.
Searchable meeting memory
Meetings contain information that becomes difficult to recover once the call ends. A transcript changes that.
Without transcription
With transcription
Meeting recording, then a structured transcript, then search.
Connect to
Spoken knowledge
Meeting transcription turns information that was trapped in audio into data other systems can retrieve.
Meeting use cases
Create searchable records of project, product, marketing, engineering, and operational discussions.
Preserve customer requirements, questions, feedback, and decisions.
Transcribe prospect conversations for CRM notes, review, coaching, and downstream analysis.
Separate interviewer and candidate or researcher and participant.
Convert qualitative discussions and focus groups into searchable text.
Maintain structured records of high-level discussions where appropriate governance and privacy controls are applied.
Turn workshops and training conversations into searchable learning material.
Transcribe conversations captured through conferencing and collaboration platforms.
Support teams that communicate across languages or naturally code-switch.
Feed transcripts into LLMs for notes, decisions, follow-up workflows, enterprise search, and knowledge systems.
Privacy & redaction
Shunya's transcription workflow includes custom keyword redaction. Custom redaction is deterministic and can be combined with transcription workflows.
Meeting transcripts may contain
Original
The account number is 483921.
Redacted
The account number is ******.
Profanity masking
****
Important. Keyword redaction is one transcript-protection tool. Organisations should still apply their own access control, retention, governance, and meeting-recording policies according to the sensitivity of the conversation.
Developers
Use the same Horizon API for meeting audio and enable the structure required by your application. Batch transcription uses POST /v1/audio/transcriptions. Choose Horizon Universal (horizon-universal), Horizon Indic (horizon-indic), or another Horizon model that fits the audio. Enable diarization with enable_diarization=true and word timestamps with response_format=verbose_json.
{
"model": "horizon-universal",
"file": "meeting.wav",
"enable_diarization": true,
"response_format": "verbose_json"
}Optional batch intelligence
For known participants
Then set enable_speaker_identification=true. For live meeting transcription, use Shunya's WebSocket streaming workflow instead of waiting for the entire meeting recording.
Basic flow
Diarization example
Build a meeting AI workflow
Once meeting speech becomes structured text, it can flow into the rest of the organisation's systems.The transcription layer converts conversation into the structured input those workflows need.
LLM / application
Business systems
One meeting
Choose the right meeting model
Global meetings
Use for broad international language coverage.
Best for:
Indian-language meetings
Use when conversations are primarily in Indian languages.
Best for:
Mixed-language meetings
Use when participants naturally switch languages during the conversation.
Best for:
Model follows the meeting
Built for production
Recording a meeting is easy. Turning it into reliable organisational data is harder. Production meeting transcription needs accurateASR, speaker separation, timing, multilingual recognition, APIs, enrichment, security controls, and deployment flexibility.
| Requirement | Horizon |
|---|---|
| Recorded Meeting Transcription | |
| Live Meeting Transcription | |
| 204+ Languages | |
| Speaker Diarization | |
| Speaker Identification | |
| Word-level Timestamps | |
| Confidence Scores | |
| Multilingual Meetings | |
| Code-switching | |
| Summarisation | |
| Intent Detection | |
| Sentiment Analysis | |
| Emotion Detection | |
| Keyterm Normalisation | |
| Custom Redaction | |
| Streaming API | |
| Batch API | |
| Enterprise Deployment |
Meeting
Questions
Meeting transcription uses speech-recognition technology to convert spoken conversations in meetings into written text.
Yes. Horizon supports both recorded and real-time speech transcription.
Yes. Speaker diarization separates transcript segments according to speaker.
Diarization works with any number of speakers and typically performs best with 2-6 distinct voices.
Yes. Registered voice profiles can be combined with diarization to identify known participants.
Shunya recommends 5-15 seconds of clean solo speech for speaker registration.
Yes. Verbose transcript output can contain each word's start time, end time, and confidence score.
Yes. Horizon supports streaming transcription with published sub-500ms first-token latency.
Yes. Completed recordings can be processed using the batch transcription API.
Yes. Summarisation is part of Shunya's optional batch intelligence layer.
Shunya's intelligence layer provides transcription and summarisation. Applications can pass that structured output into an LLM or downstream workflow to extract action items, owners, deadlines, decisions, and follow-ups.
Yes. Horizon supports 204+ languages and multilingual speech.
Horizon Code-Switch is designed for conversations where speakers naturally switch languages.
Yes. Shunya's optional batch intelligence capabilities include sentiment and emotion analysis.
Yes. Custom keyword redaction and profanity masking are available in supported transcription workflows.
On the published AMI meeting benchmark, Horizon reports 4.19% WER, compared with 9.12% for the best published baseline shown in Shunya's comparison.
No. Meeting accuracy varies with speaker overlap, microphones, accents, language, room acoustics, noise, vocabulary, and recording quality.
Yes. Horizon provides batch HTTP transcription and streaming WebSocket workflows.
Get started
Don't judge meeting transcription from our prepared conversation. Test one of yours. Upload a team meeting,customer discussion, interview, planning session, multilingual conversation, or real conference recording.See the transcript, speakers, timestamps, and meeting intelligence for yourself.