Real-world audio includes:
- Different microphones
- Spontaneous phrasing
- Local place names
- Names of people
- Interruptions
- Background noise
Horizon Indic
Turn spoken Marathi into searchable text for municipal citizen services and financial support. Shunya Labs Horizon Indic is designed for speech in Indian languages and regional varieties, with a platform for both uploaded recordings and live audio streams.
Working with Marathi audio requires attention to regional Marathi pronunciation, names, and conversational English. Test your recordings, review the transcript in the writing system your audience expects, and connect the result to the workflows that matter to your team.
माझं order अजून deliver झालेलं नाही.
माझं order अजून deliver झालेलं नाही.
Marathi speech · Horizon IndicMarathi Speech to Text
Marathi is used in Maharashtra. A single clean sample rarely captures the practical demands of a deployed speech system.
Real-world audio includes:
These details matter when a transcript feeds a support ticket, a community archive or a voice assistant.
For Marathi, pay particular attention to regional Marathi pronunciation, names, and conversational English. Select a reference script before testing, and compare the model output against transcripts checked by speakers familiar with the intended audience. A language name in a catalogue is the starting point for evaluation, not a guarantee that every dialect, script or audio channel performs identically.
Try it yourself
A prepared clip cannot show how automated speech recognition will perform on your calls, mixed-languagespeech, names, numbers, or noisy audio. Try your own Marathi recording.
Example
माझं order अजून deliver झालेलं नाही.
माझं order अजून deliver झालेलं नाही.
Another example
Use customer calls, meetings, interviews, voice notes, podcasts, sales calls, support conversations, or phone recordings.
Horizon Indic for Marathi
For supported batch configurations, consider diarization, timestamps and optional conversation intelligence. Feature availability can depend on configuration.
Streaming vs batch
Batch
Use batch transcription when you already have a Marathi interview, phone recording, podcast, lecture or field recording. The system processes the completed audio and returns text that reviewers can correct and tag. This mode is a useful starting point for archives and back-office processing because reviewers can inspect the entire recording.
Real-time
Use real-time recognition when an application needs the user's Marathi words during a conversation - for example a live caption screen, a voice-driven form or agent assistance. Shunya documents WebSocket streaming in its Horizon platform.
| Requirement | Batch | Real-time |
|---|---|---|
| Completed Marathi Recordings | - | |
| Interviews, Podcasts and Lectures | - | |
| Phone and Field Recordings | - | |
| Archives and Back-office Processing | - | |
| Live Caption Screens | - | |
| Voice-driven Forms | - | |
| Agent Assistance | - | |
| WebSocket Streaming | - |
Validate latency, audio quality and per-language recognition using the deployed service rather than assuming an aggregate benchmark applies to Marathi.
Native Marathi text
Marathi is commonly associated with Devanagari. Writing-system choices can still vary by community, context or downstream application.
Transcription
Captures the spoken language as text.
Transliteration
Changes the script.
Translation
Expresses the meaning in another language.
These are distinct operations and should be tested separately.
Before deployment, confirm how the model represents:
Keep a native-speaker-reviewed set of short recordings as a repeatable acceptance test.
Marathi translation
A typical Marathi speech-translation workflow has two clear stages: Horizon transcribes the audio into Marathi text, then Parallax translates that transcript into the supported target language.
The two outputs can be checked separately. This helps teams investigate whether a problem came from speech recognition, translation, an unexpected script or unclear source audio.
The published Parallax evaluation lists 40 BLEU for Marathi (strong published evaluation tier (still requires testing)). This is a model-level evaluation, not an accuracy guarantee for your application. Validate actual Marathi text in your required translation direction - especially proper nouns, industry-specific terms, numerals and informal speech.
Important For an English-to-Marathi experience, check output language-token and script settings first. Translation and code-switch support for a specific pair must be confirmed on the configured model.
Use cases
Voice products
Enable an in-app listening layer so people can interact with services without typing lengthy messages. The immediate opportunity for Marathi is municipal citizen services and financial support.
Operations
Connect transcripts to case management, QA review, agent assistance and downstream analytics. This is especially useful when the source material contains regional Marathi pronunciation, names, and conversational English.
Inclusion
Turn spoken questions into reviewable text for community information desks and public-facing applications. Keep human review in the workflow until your acceptance thresholds have been met.
Example
A local organisation records a Marathi service enquiry and needs a readable transcript before responding.
Illustrative workflow only: no generated native-language sentence or translation here is presented as a measured product result.
Developers
Use the Horizon ASR API for speech recognition and Parallax for text translation.
For Marathi, the editorial language identifier is mr; obtain the accepted code and model ID from the production API before putting an executable snippet on this page. Shunya's current main documentation calls the Indic model horizon-indic, while older documentation may refer to zero-indic.
Recommended configuration checklist
Questions
Marathi appears in the published Horizon Indic supported-language list. Test the current API with your own speakers, scripts, accents and recording conditions before production use.
Horizon offers both streaming and batch workflows. Verify the exact Marathi configuration and performance in the target deployment.
Marathi is associated with Devanagari. Confirm that the deployed output format matches your audience and do not assume every alternate script is automatically available.
The normal route is speech transcription followed by Parallax text translation. The published evaluation reports 40 BLEU, but real-world content and direction still require evaluation.
Use consented recordings that represent your actual speakers and tasks. Review proper nouns, regional Marathi pronunciation, names, and conversational English, and output-script consistency rather than relying only on a multilingual platform-wide benchmark.
Get started
Try real Marathi audio in the Shunya Labs playground, inspect the output, and speak with the teamabout streaming, APIs and translation support. Bring the scripts, speaker groups, audio samplesand required downstream language that your application will use.