Skip to main content
Documentation changes for the Gnani Speech APIs, newest first. Dates are the dates the documentation changed, except Prisma v3.0, which uses its release date. For Python SDK releases, see gnani-vachana on PyPI. For the other sections, see all changelogs.

​
Release notes

Detailed notes for releases that need action or migration guidance.
​
Speech-to-TextBatchNew modelAPI changes
Prisma v3.0

​
Prisma v3.0 speech-to-text model

prisma-v3.0 is a multilingual speech-to-text model for nine Indian languages and English. It can identify the spoken language by itself.Action required: none. Prisma v3.0 is opt-in and Prisma v2.5 remains the default model, so existing integrations keep working. There are no breaking changes and no deprecations in this release.

​
New models and capabilities

  • Prisma v3.0. One model for hi-IN, bn-IN, ta-IN, te-IN, kn-IN, ml-IN, mr-IN, gu-IN, pa-IN and en-IN. See Prisma v3.0 and Models.
  • Automatic language detection. Send language_code=indic-auto, or send one to four language codes as a comma-separated list such as hi-IN,en-IN. They are hints, not a lock: the transcript follows the language actually spoken. Responses include detected_language_code.
  • Two more output styles, next to verbatim and transcribe. codemix is available on REST, Batch and Realtime. transliterate writes the whole transcript in Latin letters and is available on REST and Realtime. See Output styles and Romanised output.
  • Context hints. keyterms, topic, previous_text and speaker_names can improve accuracy. On REST you can send any combination of the four. Realtime accepts keyterms only (x-keyterms).

​
API changes

  • Choose the model. REST: form field model=prisma-v3.0 on POST /stt/v3. Realtime: header x-model: prisma-v3.0 on WS /stt/v3/stream. Batch: config.model in Create Job. See Where you can use it.
  • Realtime. x-format accepts codemix and transliterate. lang_code is optional with transliterate. lang_code accepts a comma-separated list of up to four languages. See Prisma v3.0 on streaming.
  • Batch. With prisma-v3.0 the model identifies the language of each segment, and config accepts format, keyterms, topic and speaker_names. See Prisma v3.0 options.
  • New endpoint. GET /stt/v3/languages lists the languages each model supports.
  • New error types, such as INVALID_CONTEXT and MODEL_UNSUPPORTED_LANGUAGE. See the Error Reference.
  • Missing language_code on REST returns 422 with a {"detail": [...]} body instead of the usual error envelope. It is optional only with format=transliterate.

​
Improvements

  • New reference pages: Models, Prisma v2.5, Error Reference and Limits.
  • Accepted values for TTS and Voice Cloning audio_config, and for Batch bias_score and the job status filter, are now shown as tables.
  • On Realtime, segment_index starts at 1 and segment_id has the form <request_id>-seg-<N>.

​
If you switch to Prisma v3.0

Some responses differ from Prisma v2.5. Check these before you switch:
  • With a fixed hi-IN or en-IN and no format, REST returns digits (transcribe). Prisma v2.5 returned verbatim. Send format explicitly to keep your current output.
  • The model value in responses is prisma-v3.0. Update any code that checks it.
  • REST responses do not include timestamp. Use request_id instead.
  • itn_native_numerals and x-itn-native-numerals have no effect, and output.literal is not returned on Realtime. The model writes numbers itself.
The full list is in How it differs from Prisma v2.5.