> ## Documentation Index > Fetch the complete documentation index at: https://assemblyai.com/docs/llms.txt > Use this file to discover all available pages before exploring further. # Optimizing Accuracy and Latency > Tune streaming models to balance transcription accuracy and latency. ## Mode Universal-3.5 Pro Streaming workloads sit on a spectrum between two competing goals: returning transcripts as fast as possible, and returning the most accurate transcripts possible. To make this tradeoff explicit, Universal-3.5 Pro supports a **mode** connection parameter you can set when opening a streaming session. | Mode | Value | When to use | | ------------------ | -------------- | ------------------------------------------------------------------------------------------------------------------------------ | | Min latency | `min_latency` | Lowest possible time-to-text. Best when responsiveness matters more than catching every word. | | Balanced (default) | `balanced` | A middle ground between latency and accuracy. Best for voice agents and other interactive applications. | | Max accuracy | `max_accuracy` | Highest transcription accuracy. Best for note-taking, scribes, and post-call analysis where a small added delay is acceptable. | Set the `mode` connection parameter when you open the WebSocket. ```python theme={null} CONNECTION_PARAMS = { "sample_rate": 16000, "speech_model": "universal-3-5-pro", "mode": "balanced", # min_latency | max_accuracy } ``` ```python theme={null} client.connect( StreamingParameters( sample_rate=16000, speech_model="universal-3-5-pro", mode="balanced", # min_latency | max_accuracy ) ) ``` ```javascript theme={null} const CONNECTION_PARAMS = { sample_rate: 16000, speech_model: "universal-3-5-pro", mode: "balanced", // min_latency | max_accuracy }; ``` ```javascript theme={null} const transcriber = client.streaming.transcriber({ sampleRate: 16_000, speechModel: "universal-3-5-pro", mode: "balanced", // min_latency | max_accuracy }); ``` ## Language Selection By default, Universal-3.5 Pro Streaming runs in multilingual mode. Pass a `language_code` connection parameter to bias the model toward a single language. This is useful when you know the session is monolingual and want to improve language accuracy. | Model | Languages | | --------------------------- | ---------------------------------------------------------------------------------------------------------- | | Universal-3 Pro Streaming | `en`, `es`, `fr`, `de`, `it`, `pt` | | Universal-3-5 Pro Streaming | `en`, `es`, `fr`, `de`, `it`, `pt`, `tr`, `nl`, `sv`, `no`, `da`, `fi`, `hi`, `vi`, `ar`, `he`, `ja`, `zh` | Set the `language_code` connection parameter when you open the WebSocket. Omit `language_code` to keep multilingual code-switching behavior. ```python theme={null} CONNECTION_PARAMS = { "sample_rate": 16000, "speech_model": "universal-3-5-pro", "language_code": "es", } ``` ```python theme={null} client.connect( StreamingParameters( sample_rate=16000, speech_model="universal-3-5-pro", language_code="es", ) ) ``` ```javascript theme={null} const CONNECTION_PARAMS = { sample_rate: 16000, speech_model: "universal-3-5-pro", language_code: "es", }; ``` ```javascript theme={null} const transcriber = client.streaming.transcriber({ sampleRate: 16_000, speechModel: "universal-3-5-pro", languageCode: "es", }); ``` ## Advanced: Tuning turn detection parameters Beyond the `mode` parameter, you can tune individual turn detection parameters to fine-tune partial cadence and turn endpointing for your use case. The parameters differ by model. Universal-3.5 Pro Streaming uses **punctuation-based turn detection**. Turns end when terminal punctuation (`.` `?` `!`) is detected; if no punctuation is detected within `max_turn_silence`, the turn ends anyway. Each `mode` ships with its own set of defaults for these parameters. Override any of them on the connection to fine-tune further. | Parameter | Default | Description | | --------------------- | ------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `min_turn_silence` | `min_latency: 96`
`balanced: 224`
`max_accuracy: 800` | Silence (ms) before a speculative end-of-turn check fires. Lower = faster turn endings; higher = fewer entity splits on numbers and proper nouns. | | `max_turn_silence` | `min_latency: 416`
`balanced: 1536`
`max_accuracy: 1536` | Maximum silence (ms) before forcing a turn to end, regardless of punctuation. Raise it when you expect a longer pause (caller reading a credit card, address). | | `interruption_delay` | `min_latency: 0`
`balanced: 500`
`max_accuracy: 500` | Time to first partial (ms). Lower = faster TTFT for barge-in detection; higher = more confident first partials. The server adds \~300ms minimum on top. | | `continuous_partials` | `min_latency: true`
`balanced: true`
`max_accuracy: true` | When `true`, emit a partial every \~3s during continuous speech. Useful for long utterances where silence-based partials don't fire often enough. | | `vad_threshold` | `min_latency: 0.3`
`balanced: 0.2`
`max_accuracy: 0.2` | Confidence threshold (0–1) for classifying audio frames as speech. Increase for noisy environments to reduce false speech detection. | **Tuning recipe — long utterance prep** When your voice agent prompts the user for a long utterance (credit card, phone number, address), raise `min_turn_silence` mid-stream so brief pauses don't fragment the turn: ```json theme={null} { "type": "UpdateConfiguration", "min_turn_silence": 1000 } ``` After the response, restore the default: ```json theme={null} { "type": "UpdateConfiguration", "min_turn_silence": 100 } ``` See [Updating configuration mid-stream](/streaming/updating-configuration-mid-stream) for the full list of mid-stream parameters. Universal Streaming uses **confidence-based turn detection**. The model predicts when speech naturally ends; if confidence exceeds `end_of_turn_confidence_threshold` and `min_turn_silence` has passed, the turn ends. Acoustic (silence-based) detection kicks in as a fallback after `max_turn_silence`. | Parameter | Default | Description | | ---------------------------------- | --------- | ------------------------------------------------------------------------------------------------------------------------------------ | | `end_of_turn_confidence_threshold` | `0.4` | Confidence threshold for semantic end-of-turn. Higher = more confident before ending; lower = ends faster. | | `min_turn_silence` | `400` ms | Silence required before a semantic end-of-turn fires. | | `max_turn_silence` | `1280` ms | Maximum silence before forcing a turn to end via acoustic detection. | | `vad_threshold` | — | Confidence threshold (0–1) for classifying audio frames as speech. Increase for noisy environments to reduce false speech detection. | **Quick-start configurations** Aggressive — short, rapid back-and-forth (e.g., IVR replacements, order confirmations): ```javascript theme={null} const streamingConfig = { end_of_turn_confidence_threshold: 0.4, min_turn_silence: 160, max_turn_silence: 400, }; ``` Balanced — most conversational voice agents (e.g., customer support): ```javascript theme={null} const streamingConfig = { end_of_turn_confidence_threshold: 0.4, min_turn_silence: 400, max_turn_silence: 1280, }; ``` Conservative — reflective or complex speech (e.g., healthcare, sales, legal): ```javascript theme={null} const streamingConfig = { end_of_turn_confidence_threshold: 0.7, min_turn_silence: 800, max_turn_silence: 3600, }; ``` **Disabling turn detection** If you're using your own VAD or turn detection model, send a `ForceEndpoint` event to force a turn boundary: ```python theme={null} ws.send(json.dumps({"type": "ForceEndpoint"})) ``` Or set `end_of_turn_confidence_threshold` to `1` (acoustic-only fallback) or `0` (silence-only). Setting it to `0` is **not recommended** unless you have a custom turn detection model running on top — it forces a turn at every `min_turn_silence`-length pause and fragments mid-sentence thinking pauses.