# Release Notes

Here, you’ll find the latest updates, features, and fixes to our API.

For any questions or support, reach out to our team at [support@lelapa.ai](mailto:support@lelapa.ai)

## **vulavula.minor-2025.09.25-1***

- **Docs**: Re-organised API Reference documentation with subsections
- **Update**: Enhanced [uptime status page](https://lelapa.hyperping.app) to include Translate and Transcribe APIs
- **Update**: Route `zul` language Transcribe requests to new, higher accuracy model

## **vulavula.minor-2025.07.25-1**

- **Update**:  Migrated Translate  resources to the EU region
- **Update**: Observability stack upgrades
   
## **vulavula.minor-2025.06.25-1**

- **Update**: isiZulu Code Switch Improvements: We've released a new model that beta supports code-switching. This is in beta, contact us if you're interested in trying it out.

## **vulavula.minor-2025.06.18-1**

- **Update**: small tweak to improve WER on English transcriptions
- **Feature**: code-switched isiZulu support with language code cs-zul [alpha]
- **Performance**: improved latency by ~20%

## **vulavula.minor-2025.06.12#0**

- **Fix**: Transcribe Fast api connection issues.

## **vulavula.minor-2025.06.11#4**

- **Performance**: Minor performance changes to reduce latency

## **vulavula.minor-2025.05.28-1**

- **Fix**: disable expensive healthcheck process contributing to database connection issues
- **Performance**: tweak database connection pooling for improved stability
- **Update**: switched ASR model for minor improvement to WER
- **Update**: improved error handling and response when pipeline is saturated

## **vulavula.minor-2025.05.14-1**

- **Security:** Patched a minor vulnerability in the platform's front-end codebase.
- **Docs:** Updated with better explanations of HTTP Error Codes.
- **Performance:** New scaling configuration to reduce Transcribe error rates when under continuous load.

## **transcribe.sync-2025.04.16-1**

- **Update:** Improved diarisation logic for more accurate speaker identification, especially in overlapping or ambiguous segments.
- **Fix:** Addressed minor internal edge cases to better handle uncommon failure modes and improve overall robustness.
- **Feature:** Reintroduced word-level results in the response schema under `diarisation_result.words`, enabling more granular analysis. *Updated schema below.*

```json
{
  "id": "string",                      // Unique identifier
  "upload_file_size": "number",        // Size of uploaded file in bytes
  "audio_length_seconds": "number",    // Duration of audio in seconds
  "sample_rate": "number or null",     // Audio sample rate
  "channels": "number or null",        // Number of audio channels
  "frame_rate": "number or null",      // Frame rate
  "mime_type": "string or null",       // MIME type of audio file
  "language_code": "string",           // Language code (e.g., "eng")
  
  "diarisation_result": {              // Speaker identification results
    "timeline": [                      // Sequential timeline of audio segments
      {
        "start_time": "number",        // Segment start time in seconds
        "end_time": "number",          // Segment end time in seconds
        "type": "string",              // Type of segment ("speech" or "silence")
        "speaker_id": "string",        // Speaker identifier (if type is "speech")
        "text": "string"               // Transcribed text (if type is "speech")
      }
    ],
    
    "words": [                         // Detailed word-level data
      {
        "word": "string",              // Transcribed word
        "start_time": "number",        // Word start time in seconds
        "end_time": "number",          // Word end time in seconds
        "confidence": "number",        // Confidence score (0-1)
        "weight": "number or null",    // Weight value
        "word_intensity": "number or null", // Word intensity
        "best_path": "boolean or null", // Best path indicator
        "speaker_id": "string"         // Speaker identifier
      }
    ]
  },
  
  "transcription_text": "string",      // Complete transcription text
  "transcription_status": "string",    // Status (e.g., "COMPLETED")
  "error_message": "string or null",   // Error message if any
  "status_datetime": "string",         // Status timestamp
  "upload_datetime": "string",         // Upload timestamp
  "warnings": ["string"]               // Array of warning messages
}
```

## **transcribe.sync-2025.04.09-1**

- **Feature:** Introduced diarisation_result.timeline, providing a clear, turn-by-turn view of diarisation events including speech, music, and silence.
- **Update:** Refined transcription schema by removing deprecated fields from the output:
    - container_name, blob_name, customer_id, project_id, storage_url, keychain_id, batch_transcription_id
    - diarisation_result.starts, ends, words, sentences, audio_segments
- **Fix:** Resolved edge cases that were causing 500 Internal Server Errors. *While edge-case 500 errors have been patched, **occasional server errors may still occur under unhandled conditions**. We’re actively monitoring and working to resolve these.*
- **Fix:** Error messaging has been improved for clarity.
    - Invalid language codes now trigger a 400 Bad Request with the message “Invalid language code”.

## **transcribe.sync-2025.03.28-1**

- **Performance**: Adjusted baseline batch size for model inference, resulting in a major speed-up.
- **Performance**: Refactored sentence-level results algorithm providing a ~30% latency improvement.
- **Feature**: Deployed and integrated our new ASR model, improving Word Error Rate.
- **Fix**: Added Signal-to-Noise Ratio (SNR) warning for audio files that may cause inaccuracies in Diarisation results.
- **Deprecation**: Transcribe v1 endpoints have been deprecated.
- **BETA**: Music/silence detection is now integrated and available for use. Some issues are being investigated, with certain calls failing, potentially due to audio length.
- **Feature**: API release version details are now available [here](https://vulavula-services.lelapa.ai/_meta/version).
- **Note**: minor change to release naming scheme

## **transcribe.sync-2025.03.20#1**

- **Fix**: Addressed file upload issues affecting new API key owners.
- **Performance**: Upgraded Diarization model to new machines for improved performance.
- **Fix**: Resolved keychain creation bug for new platform users.

## **transcribe.sync-2025.03.19#1**

- **Fix**: Resolved HTTP 500 errors caused by SQL connection timeouts.

## **transcribe.sync-2025.03.13#1**

- **Feature**: Added `warning` messages for files with a high signal-to-noise ratio.
- **Feature**: Introduced sentence-level diarization results.
- **Fix**: Resolved issue where `null` speaker IDs appeared in some results.

## **transcribe.sync-2025.03.10#1**

- **Performance**: System stability improvements, including:
    - Enhancements to diarization model configuration.
    - Increased timeout for synchronous HTTP calls.
    - Expanded capacity to meet demand with 10x concurrency.

## **transcribe.sync-2025.03.05#1**

- **Fix**: Resolved transcoding failures for certain files; compressed WAV files are now handled correctly.
- **Fix**: Sanitized word-level diarization results by removing extraneous spaces and other formatting issues.

## **transcribe.sync-2025.03.04#1**

- **Feature**: Basic API usage reporting is now available via a [new API endpoint](https://vulavula-services.lelapa.ai/#/transcribe/get_transcribe_usage_api_v2alpha_transcribe_usage_get).
- **Feature**: Custom HTTP 503 response on the [transcribe/sync/file](https://vulavula-services.lelapa.ai/#/transcribe/sync_file_transcription_api_v2alpha_transcribe_sync_file_post) endpoint indicates when the system is not yet ready to serve requests.
- **Fix**: Resolved an issue where the incorrect Language ID was returned despite clients specifying lang_code.
- **Performance**: Various stability and performance improvements made to the Transcribe Sync pipeline.
- **Ops**: Enhanced observability signals to support troubleshooting and issue resolution.