Recorded speech features
Every recorded call passes through the same pipeline: audio preparation, transcription, optional sensitive-content masking, then the post-processing features enabled for the account.
Where recordings come from
| Source | Notes |
|---|---|
| API, base64 or URL | POST /api/v1/speech/process. URLs can be direct media links or YouTube links. |
| SFTP collection | Speechlytics collects new recordings from your SFTP server every hour. See SFTP imports. |
| Provider integrations | Microsoft Teams, Aircall, FrontSpin and 3CX. See Integrations. |
| Live calls | A finished live session is stored and processed like any other call. On accounts with polishing, it is polished automatically first. |
Supported formats
The API accepts WAV, MP3 and WMA, plus other common formats whose first stream is audio. SFTP collection picks up .wav, .mp3 and .wma files. Audio is converted to 16 kHz WAV internally.
Video files (for example MP4) are accepted from the API, URLs and SFTP when Process Video is enabled for the account; the audio track is extracted before transcription. Microsoft Teams recordings are always video and are handled automatically.
Audio preparation
Audio pre-processing is configured per account. Your Speechlytics representative can tune it for your recording sources.
| Setting | Effect |
|---|---|
| Minimum call duration | Calls shorter than the configured number of seconds are marked Failed and not transcribed. |
| Welcome message duration | Removes a fixed number of seconds from the start of every recording. See Welcome message removal. |
| Enhancement filter | Light noise filtering and gain on each channel. On by default. |
| Reduce noise / Increase volume | Additional band-pass noise reduction and volume boost for quiet or noisy sources. |
| Remove silence | Trims long silences and normalizes level before transcription. |
| Detect silent channel | When one stereo channel is silent, transcribes only the other one. |
| Check for same channels | Detects stereo files whose two channels are identical and transcribes them once. |
| Process video | Extracts audio from video sources. |
Audio content metrics (silence, music, noise and laughter per channel) are measured on every call and shown on the call's details.
Transcription models
Speechlytics transcribes with Whisper models. The model is chosen per account, trading accuracy for speed:
| Model family | Notes |
|---|---|
| Tiny, Base, Small | Fastest; suitable for clean audio or high volumes. En variants are English-only. |
| Medium | The default balance of speed and accuracy. |
| Large V1, V2, V3 | Highest accuracy for difficult audio and accents. |
| Large V3 Q8 | Large V3 quantized to 8 bits: close to Large V3 accuracy with lower resource use. |
| Large V3 Turbo | Faster Large V3 variant. |
| Small GA | A Small model fine-tuned for Irish (Gaeilge). Use it with automatic language detection. |
Live transcription uses a deployment-wide live model, independent of the account's batch model.
Language handling
Set language on the request when the spoken language is known. With 0 (Auto), Speechlytics uses the account's default transcription language if one is configured, and otherwise detects the language automatically. The detected language is stored on the call and shown in the calls list.
The Language enum lists 100 languages. Recognition quality depends on the language and the selected model.
Mono, stereo and transcribe side
| Audio | Behavior |
|---|---|
| Stereo | Left and right channels are transcribed separately. By convention the agent is on the left channel and the customer on the right. |
| Mono | A single shared Both channel. Speaker-specific metrics are not available. |
| Stereo with identical channels | Detected and transcribed once as Both. |
For stereo sources, the account's Transcribe Side setting can transcribe only one side:
| Transcribe Side | Effect |
|---|---|
| Both (default) | Both channels are transcribed. |
| Left | Only the left (agent) channel is transcribed. |
| Right | Only the right (customer) channel is transcribed. |
When only one party is transcribed (mono audio, a silent channel or a single transcribe side), analytics use prompts written for a one-party transcript, so the AI does not invent the missing side of the conversation.
Keyword matching and scoring
Configured keywords and phrases are matched after transcription. A keyword can target a channel, carry a score, be exclusive, mark a call out of scope, and contain reporting tags. Results include matched keywords, tag frequencies and the aggregate keyword score. Keywords are configured per account.
Accounts can also configure transcription filters: phrases, such as a recurring IVR announcement or recording-consent notice, whose segments are marked as filtered and excluded from transcript search.
Read Understanding call scores for the product view and Scoring model for calculation details.
AI analytics
Depending on account configuration, post-processing can add:
- a summary and structured analysis using the Sales or Customer Support profile;
- topics, sentiment, emotions, intent, business value and risk signals;
- named entities and structured call facts;
- a translated transcript;
- custom QA questions with pass/fail, evidence and extra fields;
- sensitive-content masking before any of the above;
- analysis written in the call's own language with native analytics.
After processing, a user can request transcript polishing, which corrects recognition errors with an AI model and re-runs keywords and analytics. Live calls can be polished automatically when they end.
On deployments with semantic search, processed transcripts are also indexed so calls can be found by meaning and across languages.
Agent intelligence
When call metadata identifies an agent and Intelligence is enabled, Speechlytics aggregates calls into agent scorecards, performance trends, sentiment and topic distributions and call history. See the Intelligence module.
Results and filtering
| Endpoint | Returns |
|---|---|
GET /api/v1/calls/{id}/status | One call, with the complete transcript, translation and analytics. |
GET /api/v1/calls | Calls filtered by date, filename, topic, sentiment, tag, content or cluster, newest first. Up to 100 per page. Content uses semantic search when the deployment enables it. |
GET /api/v1/favorites | The account's favorite calls, with the same filters. |
GET /api/v1/analytics | Paginated audio and score analytics for a date range. |
GET /api/v1/topics, GET /api/v1/sentiments | Account-level topic and sentiment distributions. |
GET /api/v1/intelligence | Agent-level analytics. |
List endpoints return the stereo leftChannel and rightChannel transcript segments only. Use GET /api/v1/calls/{id}/status to read mono (bothChannels) transcripts and translations.
See the API reference for exact query parameters.