.png)
Azure AI Speech Integration Guide
Integrate Azure AI Speech with enterprise systems through REST APIs, Speech SDKs, asynchronous transcription, audio files, and secure authentication.
Azure AI Speech integration options at a glance
Azure AI Speech provides REST APIs and Speech SDKs for speech-to-text, batch transcription, text-to-speech, speech translation, custom Speech models, and related capabilities. Martini can consume the REST APIs, securely manage subscription keys or bearer tokens, submit audio for asynchronous batch processing, poll job status, retrieve result files, and map outputs into business applications or databases. Audio inputs and generated files can be handled through documented storage and file patterns. Selected asynchronous operations may support callback-style notifications, but polling is the broadly applicable approach. Persistent real-time streaming may require a custom JVM-compatible component using an Azure Speech SDK.
| Integration point | Supported by Azure AI Speech? | Common use cases | How Martini supports it |
|---|---|---|---|
| REST APIs | Yes | Speech-to-text, text-to-speech, token acquisition, batch transcription, and custom Speech operations use documented REST endpoints. | Martini can consume the Azure AI Speech REST APIs, manage request headers and payloads, map responses, and orchestrate multi-step workflows. |
| Speech SDKs | Yes | SDKs support language-specific recognition, synthesis, and streaming scenarios across supported programming languages. | Martini can use REST directly or invoke custom JVM-compatible logic when an SDK-specific capability is required. |
| Bulk / async / batch APIs | Yes | Batch transcription processes one or more audio files asynchronously and exposes job status and result artifacts. | Martini can submit jobs, persist job identifiers, poll status, retrieve result files, and route completed transcripts. |
| File / attachment APIs | Limited | Audio files can be supplied for transcription and batch jobs can produce result files; this is not a general-purpose attachment API. | Martini can handle file references and result artifacts, validate media metadata, and forward outputs to storage or downstream applications. |
| Webhooks / outbound callbacks | Limited | Callback or notification behavior is available for selected asynchronous operations and API surfaces, rather than every Speech event. | Martini can receive a documented callback, but should use scheduled polling when the selected operation does not provide a reliable notification mechanism. |
| Authentication | Yes | Speech APIs support subscription keys, short-lived bearer tokens, and supported Microsoft Entra ID or managed identity configurations depending on the operation. | Martini can store credentials in secrets or environment configuration and apply the required authentication to outbound API calls. |
| Streaming protocols | Limited | Real-time recognition and synthesis use streaming connections and Speech SDK transport mechanisms rather than ordinary request-response REST calls. | Martini can orchestrate surrounding APIs, while persistent bidirectional streaming may require a custom JVM-compatible component or application service. |
| Database / analytics access | Not confirmed | Azure AI Speech does not expose a documented relational database interface; processed results can be sent to a database by Martini. | Martini can map Speech results into SQL operations or downstream analytics ingestion paths. |
How Azure AI Speech exposes data and business events
Azure AI Speech REST APIs
Azure AI Speech exposes REST APIs for speech recognition, text-to-speech, token acquisition, batch transcription, and related Speech operations. REST is the most direct integration surface for server-to-server workflows.
Martini implementation pattern
Martini implementation pattern: Martini uses a configured REST-consuming workflow to authenticate, construct the Speech request, submit audio or text, validate the response, and map the result into a canonical model or downstream API.
Implementation sequence
Batch transcription
Batch transcription processes one or more audio files asynchronously. A client submits a job, checks its status, and retrieves result files after completion.
Martini implementation pattern
Martini implementation pattern: Martini submits the batch request, persists the Azure job identifier and source correlation key, then uses a scheduled or queue-oriented workflow to poll status and retrieve results without creating duplicate jobs.
Implementation sequence
Audio and result files
Azure AI Speech accepts audio input and can return audio output or transcription result files, depending on the selected operation. Durable storage such as Azure Blob Storage may be used where documented by the chosen API.
Martini implementation pattern
Martini implementation pattern: Martini manages file references rather than embedding large binaries in ordinary messages when appropriate, validates access and media properties, and routes result artifacts to approved storage or enterprise applications.
Implementation sequence
Selected callbacks and notifications
Azure Speech supports callback or webhook-style notifications for selected asynchronous operations and API surfaces. This behavior does not apply to every Speech event or every real-time recognition result.
Martini implementation pattern
Martini implementation pattern: Martini can expose an API for a documented callback, validate the notification, and reconcile it with persisted job state. Scheduled polling remains the fallback for batch operations without a suitable callback.
Implementation sequence
Speech SDK and streaming scenarios
Real-time recognition and synthesis use streaming connections and Speech SDK transport mechanisms. These are materially different from ordinary REST request-response calls.
Martini implementation pattern
Martini implementation pattern: Martini can orchestrate the surrounding request, authentication, storage, and downstream processing, while a custom JVM-compatible component can encapsulate the Azure Speech SDK when persistent bidirectional streaming is required.
Implementation sequence
Common Azure AI Speech integration patterns
Pattern 1: Transcribe contact-center audio into case records
When to use this pattern
Use this pattern when recordings from a contact center or collaboration platform must be transcribed and attached to customer-service records. The workflow should support asynchronous processing, speaker metadata where available, and reconciliation of retries against the original interaction.
Integration direction
Example Mapping
| Azure AI Speech Field | Canonical Field | Target Field |
|---|---|---|
| sourceInteractionId | interactionId | correlation_id |
| transcription.phrases | transcriptText | work_notes |
| transcription.timestamps | transcriptSegments | transcript_metadata |
| job.status | processingStatus | u_speech_status |
Martini implementation pattern
A Martini API or scheduled workflow receives an audio reference, validates the media and source identifier, submits a batch transcription job, and stores the job state. A follow-up workflow polls for completion, maps phrases and timestamps, applies rules for missing or low-quality output, and updates ServiceNow. Stable interaction identifiers and Azure job IDs prevent duplicate case updates; transient failures use backoff and retry.
Martini capabilities used
- APIs
- workflows
- API consumption
- data mapping
- business rules
- scheduled execution
- error handling
Pattern 2: Convert voice intake into structured business requests
When to use this pattern
Use this pattern when an application accepts an audio recording for a service, maintenance, or operational request and downstream systems require structured fields rather than unprocessed speech output.
Integration direction
Example Mapping
| Azure AI Speech Field | Canonical Field | Target Field |
|---|---|---|
| audioReference | mediaLocation | documentReference |
| transcription.text | requestDescription | longText |
| transcription.language | locale | language |
| sourceRequestId | requestId | externalReference |
Martini implementation pattern
Martini receives the request through an exposed API, validates the audio location and required metadata, invokes Azure AI Speech, and applies extraction and validation rules to the transcript. Only requests meeting required-field and confidence policies are sent to SAP S/4HANA; incomplete results are routed for review. The workflow records correlation identifiers and retries transient API failures without resubmitting an already accepted job.
Martini capabilities used
- API exposure
- API consumption
- workflows
- validation
- mapping and transformation
- business rules
- retry handling
Pattern 3: Generate audio notifications from enterprise text
When to use this pattern
Use this pattern when a business application needs Azure AI Speech to convert approved text into audio for a notification, call-center, or accessibility workflow.
Integration direction
Example Mapping
| Azure AI Speech Field | Canonical Field | Target Field |
|---|---|---|
| text | messageText | input_text |
| voice | voiceName | voice |
| language | locale | language |
| outputFormat | audioFormat | content_type |
Martini implementation pattern
A Martini API validates text length, locale, voice, and output-format parameters before calling the Speech synthesis endpoint. The workflow returns the audio for smaller payloads or stores it in an approved location for larger outputs, then forwards a controlled reference to the target application. Invalid parameters are rejected before the Azure call and transient failures are retried according to the operation's safety rules.
Martini capabilities used
- API exposure
- API consumption
- validation
- data mapping
- workflow orchestration
- secure configuration
- error handling
Pattern 4: Manage custom Speech model operations
When to use this pattern
Use this pattern when domain terminology, pronunciation data, or approved training datasets must be registered and incorporated into later transcription workflows. Exact operations and permissions depend on the selected API version and region.
Integration direction
Example Mapping
| Azure AI Speech Field | Canonical Field | Target Field |
|---|---|---|
| datasetReference | trainingDataLocation | dataset_uri |
| modelName | modelIdentifier | custom_model_id |
| operation.status | modelStatus | deployment_status |
| approvedLocale | locale | speech_locale |
Martini implementation pattern
A scheduled Martini workflow detects approved dataset changes, calls the documented custom Speech endpoints, tracks operation and model identifiers, and records the resulting state. Business rules prevent unapproved models from being selected by production transcription workflows. Authentication failures, invalid datasets, region mismatches, and transient service errors are separated for remediation or retry.
Martini capabilities used
- scheduled workflows
- API consumption
- data mapping
- business rules
- state tracking
- secrets management
- error handling
Applications commonly integrated with Azure AI Speech
Azure AI Speech can be used alongside recording platforms, customer-service applications, collaboration tools, document repositories, and enterprise business systems. The exact access path depends on the source application, its APIs, tenant configuration, and the selected Speech operation. Martini can coordinate those endpoints, normalize speech results, and route the output to downstream systems.
| Application | Scenario | Direction | Martini Pattern |
|---|---|---|---|
| Microsoft Teams | Transcribe meeting, call, or customer-interaction audio and route transcripts for retention, search, or follow-up processing. | Microsoft Teams → Martini → Azure AI Speech → Microsoft Teams | Martini receives an approved recording reference, submits or retrieves audio through the configured Microsoft APIs, starts Azure AI Speech processing, polls asynchronous jobs when required, and writes normalized transcript metadata back to Microsoft 365 or another target. |
| Salesforce | Attach call transcripts to Leads, Contacts, Accounts, Cases, or Opportunities to support sales and service analysis. | Salesforce → Martini → Azure AI Speech → Salesforce | A Martini workflow accepts an audio reference, validates language and media metadata, calls Azure AI Speech, maps phrases, timestamps, and speaker data, and writes the completed transcript to the appropriate Salesforce object with duplicate protection. |
| ServiceNow | Convert service-call recordings into incident notes, case comments, or searchable knowledge content. | ServiceNow → Martini → Azure AI Speech → ServiceNow | Martini receives a ServiceNow or contact-center request, submits audio to Azure AI Speech, monitors the batch job, applies validation and routing rules, and updates the related ServiceNow record with transcript content and processing status. |
| Microsoft Dynamics 365 | Enrich customer-service or sales records with transcriptions and extracted action items. | Microsoft Dynamics 365 → Martini → Azure AI Speech → Microsoft Dynamics 365 | Martini orchestrates the audio reference, Speech request, asynchronous status checks, transcript transformation, and Dynamics 365 update while preserving the source interaction identifier for reconciliation. |
| SharePoint | Store audio and generated transcript files in document libraries with metadata and retention controls. | SharePoint → Martini → Azure AI Speech → SharePoint | A Martini workflow retrieves or receives a controlled SharePoint file reference, submits the audio to Azure AI Speech, retrieves the result artifact, and writes the transcript and processing metadata to the appropriate library location. |
| Power BI | Provide normalized transcription metrics, speaker statistics, processing KPIs, or downstream classifications for reporting. | Azure AI Speech → Martini → Power BI | Martini extracts relevant Speech result fields, normalizes timestamps and processing states, stores the data in an approved analytical destination, and exposes or forwards the resulting dataset through the selected Power BI ingestion path. |
| SAP S/4HANA | Use voice-derived data to initiate or enrich service, maintenance, or operational processes. | Audio source → Martini → Azure AI Speech → SAP S/4HANA | Martini submits approved audio for transcription, applies business rules to extract structured values, validates required fields, and calls the relevant SAP API while recording the source interaction and Azure job identifiers. |
| Zendesk | Add call transcripts to tickets and use extracted text to support triage, categorization, or quality review. | Zendesk → Martini → Azure AI Speech → Zendesk | Martini receives a ticket-linked audio reference, submits it for Speech processing, transforms the completed transcript, and updates the Zendesk ticket or attachment workflow with retry and duplicate controls. |
How to build a Azure AI Speech integration in Martini
Objective
Configure the Azure region, Speech endpoint, API version, and authentication method for each environment without embedding credentials in workflow payloads.
Instructions in Martini
- Store subscription keys, bearer-token settings, or supported Microsoft Entra ID configuration in Martini secrets or environment configuration.
- Keep development, testing, and production Speech resources separated where appropriate.
- Confirm that the selected authentication method is supported by the Speech operation and region.
Objective
Select the trigger that matches the processing model: an exposed API for incoming requests, a schedule for batch status checks, or a documented callback for selected asynchronous operations.
Instructions in Martini
- Use an API trigger for audio locations, synthesis requests, or external processing requests.
- Use a scheduler or queue-oriented workflow for batch transcription polling.
- Use a callback endpoint only when the selected Azure operation explicitly documents notification support.
Objective
Obtain the audio, text, job state, or result artifact required by the Speech operation while controlling payload size and storage access.
Instructions in Martini
- Validate audio references, storage permissions, format, duration, language, and channel information.
- Retrieve current job status and result files through documented Azure endpoints.
- Avoid placing large binary content or sensitive media URLs in ordinary logs and messages.
Objective
Coordinate the Azure request, asynchronous state transitions, result retrieval, and downstream writes as a durable Martini workflow.
Instructions in Martini
- Persist the source identifier, Azure job identifier, processing state, and correlation information.
- Separate submission, status polling, result retrieval, and downstream delivery where asynchronous processing requires it.
- Use conditional routing for succeeded, failed, canceled, partial, and retryable states.
Objective
Convert Azure Speech request and response structures into canonical enterprise models while preserving useful metadata.
Instructions in Martini
- Map phrases, timestamps, confidence-related information, speakers, channels, locale, and result-file references where available.
- Tolerate optional speaker or channel metadata because not every operation returns the same fields.
- Normalize audio output references and transcript segments before sending them to target systems.
Objective
Validate business requirements and control duplicate processing before data is written to downstream applications.
Instructions in Martini
- Use a stable interaction ID, source recording ID, checksum, or upstream event ID for idempotency.
- Reject unsupported audio or incomplete structured requests before invoking downstream business APIs.
- Route low-quality, incomplete, or policy-sensitive results for review rather than silently accepting them.
Common Azure AI Speech data objects used in integrations
| Object | Typical Use | Common target systems | Martini handling |
|---|---|---|---|
| Speech resource | Defines the Azure endpoint, region, credentials, quotas, and billing scope used for Speech operations. | Martini configuration, Azure administration, monitoring stores | Martini keeps the endpoint, region, API version, and credentials in environment configuration or secrets rather than workflow payloads. |
| Audio input | Provides the stream, request body, file, or externally accessible media location submitted for recognition or transcription. | Azure AI Speech, Azure Blob Storage, SharePoint, Microsoft Teams, contact-center platforms | Martini validates format, size, duration, language, channel information, and access permissions before submitting the audio or its documented reference. |
| Transcription | Contains recognized phrases, timestamps, confidence-related information, and speaker or channel metadata where supported. | Salesforce, ServiceNow, Microsoft Dynamics 365, Zendesk, databases, analytics stores | Martini maps required fields into a canonical transcript model, tolerates optional speaker metadata, and applies business rules before writing downstream. |
| Batch transcription job | Represents an asynchronous request that processes one or more audio files and exposes status and result artifacts. | Martini workflow state, SQL databases, queues, operational applications | Martini stores the job identifier and source correlation key, polls documented status fields, prevents duplicate submissions, and retrieves results after completion. |
| Speech synthesis request and audio output | Combines text, language, voice, speaking style, and output format to produce synthesized audio. | Notification services, call-center platforms, storage systems, enterprise applications | Martini validates request parameters, calls the synthesis endpoint, and returns or stores the generated audio according to payload-size and retention requirements. |
| Custom Speech model | Represents a domain-specific recognition model trained or adapted with supported datasets such as audio, transcripts, or pronunciation data. | Azure AI Speech, model registries, transcription workflows | Martini can orchestrate documented dataset and model operations, record model identifiers, and select approved models in later transcription workflows subject to API and region availability. |
Authentication and security considerations
Credential protection
Azure AI Speech supports subscription keys, short-lived bearer tokens, and supported Microsoft Entra ID or managed identity configurations depending on the selected operation and deployment. Store these values in Martini secrets or environment configuration rather than workflow payloads, URLs, client applications, or ordinary logs.
Endpoint and resource controls
Keep the Azure region, Speech resource, API version, and authentication configuration aligned. Use separate resources or credentials for development, testing, and production where appropriate, and grant only the permissions required by the selected Speech operation.
Speech data handling
- Protect audio, transcripts, generated audio, and temporary media URLs as potentially sensitive data.
- Apply retention, deletion, residency, encryption, and access-control policies appropriate to the content.
- Use private storage and short-lived access URLs when external file references are required.
- Avoid exposing credentials or sensitive speech payloads through client-facing responses and logs.
Operational considerations for Azure AI Speech integrations
Quotas and retries
Speech resources can impose quotas and throttling. Limit concurrent submissions and use exponential backoff for transient 429 and 5xx responses. Separate interactive and batch workloads where the Azure architecture and governance model permit it.
Asynchronous state
Batch transcription requires durable job state. Store the source identifier, Azure job ID, status, and result reference, and distinguish running, succeeded, failed, and canceled states. Reconcile existing jobs before retrying submission.
Media and schema validation
- Validate codec, container, sampling rate, channel layout, size, duration, language, and storage accessibility.
- Do not assume every transcript includes speaker, channel, confidence, or timestamp metadata.
- Follow documented pagination fields, continuation links, result manifests, and downloadable artifacts.
- Keep API versions and regional settings configurable and test changes to response structures.
Monitoring and testing
Record correlation identifiers and Azure request identifiers where available, while excluding sensitive content from logs. Test authentication failures, invalid media, quota responses, inaccessible storage, partial result files, duplicate submissions, and downstream write failures.
Why use Martini instead of scripts or point-to-point integrations?
Centralized orchestration
Martini coordinates Azure AI Speech calls, asynchronous job monitoring, file handling, downstream APIs, databases, and business rules in maintainable workflows rather than scattering behavior across scripts.
Reusable integration assets
Teams can expose a consistent API façade, reuse authentication and transformation logic, and standardize transcript, audio, and job-state models across applications such as customer service, collaboration, and enterprise operations.
Reliability and control
- Apply validation, idempotency, retries, conditional routing, and durable status tracking.
- Keep credentials and environment-specific endpoints outside application code.
- Separate provider-specific Speech behavior from downstream system mappings.
- Monitor and troubleshoot integration workflows through centralized operational controls.
Frequently asked questions
Azure AI Speech can be integrated through its REST APIs, Speech SDKs, asynchronous batch transcription endpoints, audio and result-file workflows, and supported authentication methods. Enterprise workflows can submit audio or text, monitor processing, retrieve transcripts or synthesized audio, and route results to applications, databases, storage, or messaging systems.
Yes. Martini can consume Azure AI Speech REST APIs, manage subscription-key or token-based authentication, orchestrate batch transcription, poll job status, retrieve results, transform Speech data, and expose APIs for upstream applications. Specialized persistent streaming scenarios may require custom JVM-compatible code using a Speech SDK.
No. A dedicated Azure AI Speech connector is not required. Martini can integrate using Azure AI Speech's native REST APIs, authentication methods, batch endpoints, documented file or result interfaces, and selected callback mechanisms.
Lonti does not charge an additional per-connector or per-vendor fee to integrate Azure AI Speech with Martini. Integrations are subject to the provisioned capacity of the Martini environment. Separate costs may apply from Microsoft Azure, infrastructure providers, storage services, or other third-party systems based on subscription, usage, and deployment model.
REST APIs are the primary choice for speech recognition, synthesis, token acquisition, batch transcription, and related operations. Speech SDKs are relevant for language-specific or streaming capabilities, while batch APIs are appropriate for asynchronous file processing. GraphQL and SOAP APIs are not confirmed for Azure AI Speech.
Callback or webhook-style notifications are available for selected asynchronous operations and API surfaces, but they should not be assumed for every Speech event or real-time recognition result. Martini can receive a documented callback, while scheduled polling remains the broadly applicable pattern for batch transcription.
Batch transcription is synchronized by persisting the source recording identifier, Azure job identifier, processing state, and result reference. Martini can poll documented status fields, follow result manifests or pagination where required, and reconcile retries against stable identifiers so that the same audio is not submitted or written downstream more than once.
Yes. Martini can expose a controlled REST API that accepts an audio location, text-to-speech request, metadata, or batch request and then invokes Azure AI Speech behind the API. It can validate inputs, hide Speech credentials, apply business rules, standardize responses, and route failures without exposing provider-specific details to every client.
Related Martini documentation
Workflows
Data
Connect Azure AI Speech with Martini
Use Martini to orchestrate Azure AI Speech REST APIs, secure asynchronous transcription workflows, speech synthesis, file handling, and downstream enterprise integrations.