Upload
Real podcast formats
- An interview podcast
- Solo episode
- Panel discussion
- Video podcast audio
- Business podcast
- Educational episode
- News discussion
- Multilingual conversation
Horizon · Podcast Transcription
Turn podcast episodes into accurate, speaker-aware transcripts for search, show notes, captions, summaries, editing, archives, accessibility, and AI workflows. Use Shunya Horizon to transcribe podcast audio across 204+ languages - with speaker labels, word-level timestamps, multilingual recognition, and optional post-processing intelligence.
Welcome back. Today we're talking about how AI is changing customer support.
Speaker diarization · HorizonPodcast Transcription
Podcasts contain hours of valuable spoken information - interviews, stories, expert discussions, product conversations, news analysis, and educational content. But once an episode is published, that information is difficult to search unless the speech becomes text. Horizon turns podcast audio into structured transcripts.
204+
Languages across the current Horizon platform
3.10%
Composite WER across 8 OpenASR benchmarks
4.99%
Published WER on GigaSpeech - podcasts / YouTube-style speech
47%
Relative improvement over the published GigaSpeech baseline comparison
| Capability | Horizon · Podcast Transcription |
|---|---|
| Long-form Audio | |
| Hosts + Guests | |
| Speaker Diarization | |
| Speaker Identification | |
| Word-level Timestamps | |
| Confidence Scores | |
| Multilingual Podcasts | |
| Code-switching | |
| Podcast Summaries | |
| Keyterm Normalisation | |
| Redaction | |
| API Integration |
Try it yourself
Podcast transcription should be evaluated on actual long-form speech.
Upload
Example
HOST Welcome back. Today we're talking about how AI is changing customer support.
GUEST I think the biggest change is that companies can now respond in multiple languages without rebuilding the entire support stack.
HOST So where does speech recognition fit into that?
Model
Horizon
Enable
Transcript structure
After transcription
Optionally add
Podcast benchmark
One of the most relevant benchmarks in Shunya's published evaluation is GigaSpeech - a large-scale speech dataset containing podcasts, YouTube-style audio, and other real-world spoken media.
4.99%
Horizon WER on GigaSpeech
47%
Published relative improvement
| Model | GigaSpeech WER |
|---|---|
| Horizon | 4.99% |
| NVIDIA Canary-Qwen | 9.43% |
| Best published baseline shown by Shunya | 9.43% |
Why this matters
Read this fairly
It is not a guarantee that every podcast episode will achieve 4.99% WER. Actual results depend on:
Benchmark evidence first. Your own podcast second.
Hosts & guests
A podcast transcript becomes confusing when every speaker is merged into one paragraph.
Without diarization
Welcome to the show. Thanks for inviting me. Let's start with your latest research.
With diarization
SPEAKER_00 Welcome to the show.
SPEAKER_01 Thanks for inviting me.
SPEAKER_00 Let's start with your latest research.
Speaker diarization
Use for
Host + guest
Two-host shows
Panel podcasts
Interview series
Roundtables
Research discussions
Identify recurring hosts
If the same host appears in every episode, anonymous labels are less useful. Speaker identification can resolve registered voices into known identities.
Before
Welcome back to the podcast.
After
Welcome back to the podcast.
Workflow
Unknown guest voices can remain SPEAKER_01 or another anonymous label.
Best use
Recurring podcast hosts
Regular co-hosts
Recurring experts
Internal company podcast presenters
Multiple speakers
Many shows include more than two voices. Shunya's current diarization documentation says the feature
works with any number of speakers, with best results typically around 2-6 distinct speakers.
2-6
Distinct speakers where diarization generally performs strongest
Many shows include
Podcast audio
Diarization gets harder with
Speaker overlap
Frequent interruptions
Similar voices
Heavy echo
Background music
Poor remote audio
Good recording quality still matters.
Word-level timestamps
Podcast transcripts become much more useful when each word remains connected to the audio timeline.
Horizon verbose output can include start time, end time, and confidence score.
Search
The transcript becomes an index into the podcast.
Search inside podcasts
Without transcription
Typical workflow
With transcription
Workflow
Build search across
One episode
One season
Hundreds of episodes
An entire media archive
Spoken content
A podcast archive can become a text-searchable information library.
Podcast summaries
A transcript preserves the complete conversation. A summary helps users understand the episode more quickly. Shunya's batch intelligence layer includes summarisation.
Podcast
Use for
Episode descriptions
Internal review
Research
Archives
Knowledge bases
Content discovery
Important
Show notes
Chapter titles
Action items
SEO descriptions
Social posts
Those can be generated downstream by an LLM using the transcript and summary as inputs.
Transcribe first
Show notes workflow
Podcast transcription can become the foundation of a larger publishing workflow.
Workflow
Generate
Show notes
Episode description
Key takeaways
Guest highlights
Topic list
Quote candidates
Horizon provides transcription and supported intelligence. Show-note writing itself belongs in the downstream content-generation layer. This separation keeps the original transcript available as the source of truth.
Podcast chapters
Long podcasts often cover several topics. Word and segment timestamps give applications the raw structure needed to create navigable chapters.
Transcript
Topic A begins at 00:00
Topic B begins around 12:34
Topic C begins around 28:17
Chapter workflow
Chapter list
00:00 Introduction
12:34 Voice AI
28:17 Multilingual Support
43:50 Deployment
Horizon provides the speech and timing foundation. Chapter generation can be added downstream.
Clips & quotes
Podcast teams often need short clips for social, promotions, highlights, newsletters, and video snippets.
Teams need clips for
Social media
Promotions
Highlights
Newsletters
Video snippets
Transcript search · find
Interesting claim
Strong quote
Important answer
Product mention
Guest insight
Example
Text-based discovery helps editors find moments before they start scrubbing through the waveform.
Multilingual podcasts
Global podcasts and Indian podcasts frequently switch languages within the same episode.
HOST: So what happened after the launch?
GUEST: Launch के बाद हमें बहुत अच्छा response मिला, especially tier-two cities से.
One episode
Model options
Global podcasts.
Model options
Indian-language podcasts.
Model options
Mixed-language conversations.
Use for
Podcasts across languages
Horizon supports broad multilingual transcription - different shows, different languages, same transcription architecture.
Global speech
Indian speech
Searchable multilingual archive
From podcast to intelligence
For supported batch workflows, Shunya's intelligence layer can enrich a podcast transcript.
Who spoke when?
Which registered host was speaking?
What dominant emotion appears at different points?
What conversational intent appears in the transcript?
What broad sentiment is present?
What happened across the episode?
Normalise important names and terminology.
Connect text back to audio.
Mask configured sensitive terms.
Podcast
Keyterm normalisation
Podcasts frequently contain unusual vocabulary. Generic transcription may render important termsinconsistently - making search harder even when the transcript reads well.
Podcast vocabulary
Company names
People
Product names
Research terminology
Technical acronyms
Industry language
Example
Use for
Company names
Guest names
Product names
Technical terms
Scientific vocabulary
Industry acronyms
Media editing workflow
Transcription can make audio editing faster.
Workflow
Find
Topic
Quote
Mistake
Repeat
Sponsor segment
Section transition
Media use case
Long interviews
Video podcasts
Panel recordings
Documentary-style shows
Research-based episodes
The transcript becomes a map of the audio.
Audio & video podcasts
Horizon batch transcription supports common audio files. Video podcast files can also be processed through Shunya's supported video workflow.
Audio podcasts
WAV
MP3
M4A
OGG
FLAC
WebM
Video podcast
MP4
MOV
MKV
AVI
WebM
Video podcast
The ASR model transcribes what was spoken. Visual content still requires a separate vision or OCR layer if needed.
Developers
Send podcast audio or supported video to the batch transcription endpoint. Choose the model according to language - for example horizon-universal for global-language podcasts or horizon-indic for supported Indian-language episodes. Enable diarization and verbose JSON for word-level timing.
curl -X POST https://asrv2prod.shunyalabs.ai/v1/audio/transcriptions \
-H "Authorization: Bearer $ACCESS_TOKEN" \
-F "file=@episode.mp3" \
-F "model=horizon-universal" \
-F "enable_diarization=true" \
-F "response_format=verbose_json"Optional batch intelligence
Enrich completed transcripts
Basic flow
Choose the right podcast model
Global podcasts
Use for international-language speech.
Best for:
Indian-language podcasts
Use for supported Indian languages and regional speech.
Best for:
Mixed-language podcasts
Use when hosts and guests naturally move between languages.
Best for:
Model follows the conversation
Built for production
A useful podcast transcription workflow needs more than plain text. It needs speaker structure, timestamps, multilingual recognition,media-format support, APIs, summaries, terminology handling, and connections to downstream content and search systems.
| Requirement | Horizon · Podcast Transcription |
|---|---|
| GigaSpeech Benchmark (4.99% WER) | |
| Audio Podcasts | |
| Video Podcasts | |
| 204+ Languages | |
| Speaker Diarization | |
| Speaker Identification | |
| Word-level Timestamps | |
| Confidence Scores | |
| Multilingual Recognition | |
| Code-switching | |
| Summarisation | |
| Intent / Sentiment / Emotion | |
| Keyterm Normalisation | |
| Redaction | |
| Audio + Video Input | |
| API Integration | |
| Enterprise Deployment |
Production flow
Questions
Podcast transcription uses speech-recognition technology to convert spoken podcast audio into written text.
Yes. Horizon supports batch transcription of long-form recorded speech and common media formats.
Yes. Horizon reports 4.99% WER on GigaSpeech, which Shunya's current capability documentation identifies as a Podcasts / YouTube benchmark.
No. Results vary with speakers, microphones, language, crosstalk, background music, vocabulary, and recording quality.
Yes. Speaker diarization assigns speaker labels to transcript segments.
Shunya currently documents diarization as supporting any number of speakers, with best results generally around 2-6 distinct voices.
Yes. Registered speaker profiles can be combined with speaker identification to resolve known voices.
Yes. Verbose output includes word-level start time, end time, and confidence information.
The transcript can be indexed and searched by your application. Word timestamps can then connect the matching text back to the relevant point in the episode.
Yes. Summarisation is part of Shunya's current optional batch intelligence layer.
Shunya supports transcription and summarisation. A downstream LLM or content workflow can use those outputs to generate show notes, descriptions, key takeaways, and other publishing assets.
Horizon provides transcript and timestamp information. Chapter titles and topic boundaries can then be generated by an LLM or selected by an editor.
Yes. Word-level timestamps make it possible for an application or editor to connect a quote in the transcript to its location in the source audio.
Yes. Horizon provides broad multilingual speech recognition, while specialised Horizon Indic and Horizon Code-Switch models support additional Indian-language and mixed-language workflows.
Yes. Shunya's media workflow supports common video containers by extracting their audio track before transcription.
Yes. Shunya's current batch intelligence layer includes sentiment and emotion features, among other enrichments.
Yes. Keyterm normalisation can help standardise important names, product terms, and domain vocabulary.
Yes. Custom redaction and profanity masking are available in supported batch workflows.
Yes. Horizon provides batch transcription through the speech-to-text API.
Get Started
Don't judge podcast transcription from our prepared clip. Test one of yours. Upload an interview, panel,solo episode, multilingual conversation, video podcast, business show, or real long-form recording.See the transcript, speakers, timestamps, and structured output for yourself.