0.10.1

Release 0.10.1

Added

• Flux text-to-speech (/v2/speak) support in the new speak::flux module:
  • Batch (REST): Speak::flux_speak_to_file and Speak::flux_speak_to_stream synthesize a complete block of text with POST /v2/speak. Query options cover model (required, e.g. Model::FluxHaleyEn), encoding, sample_rate, speed, expressivity, mip_opt_out, tag, and the REST-only container, bit_rate, callback, callback_method, and priority.
  • Streaming (WebSocket): Speak::flux_request(options).handle() connects to the /v2/speak websocket and returns a FluxSpeakHandle with speak, flush, interrupt, configure_speed, close, and receive. Server events surface as FluxSpeakResponse: binary Audio frames plus Connected, SpeechStarted, Flushed, SpeechMetadata, SpeechInterrupted, SessionMetadata, ConfigureSuccess/ConfigureFailure, Warning, FatalError, and a forward-compatible Unknown.
  • The websocket rejects REST-only options and the REST-only compressed encodings (mp3, opus, flac, aac) up front with the new DeepgramError::InvalidOptions variant; custom encodings are left for the server to validate.
  • New examples: flux_tts_batch and flux_tts_websocket under examples/speak/flux/.
• FluxHandle::force_end_turn() sends the ForceEndTurn control message to end the current turn immediately on an external signal (push-to-talk release, DTMF tone, UI event, or an external endpointing stack). Sent while no turn is active, it is silently ignored.
• FluxResponse::TurnInfo now exposes trigger (Option<TurnTrigger>), the cause of a turn ending, reported on EndOfTurn events: Model (native end-of-turn detection), Manual (a ForceEndTurn was sent), or Timeout (eot_timeout_ms elapsed). TurnTrigger is an open enum — unrecognized values deserialize as TurnTrigger::Unknown(String), preserving the original wire value so it survives re-serialization.
• FluxWord now exposes optional per-word start and end timings, in seconds.
• New example: flux_force_end_turn under examples/transcription/flux/.
• Opt-in connect diagnostics behind the new connect-diagnostics cargo feature, for live transcription (/v1/listen) WebSocket connections (other WebSocket surfaces are not covered yet). Configuring a sink via WebsocketBuilder::diagnostics makes the SDK establish the streaming connection in four individually timed phases (DNS, TCP connect, TLS handshake, WebSocket upgrade) and emit one diagnostics::ConnectRecord per connect attempt. New example: connect_diagnostics.
  • Records are emitted for every attempt — including attempts cancelled by a caller-side tokio::time::timeout, and failed upgrades, which capture the dg-request-id and dg-error response headers.
  • Records serialize to flat JSON for JSONL pipelines (schema_version: 1, additive-only). The recorded URL is reduced to scheme, host, port, and path — userinfo and query parameters are never persisted.
  • TLS note: with the feature enabled, both the timed and the untimed /v1/listen connect paths are handed one explicit rustls connector (webpki trust roots, crate-default provider), so downstream feature unification — e.g. a consumer also enabling tokio-tungstenite/native-tls — cannot make the two paths select different TLS providers. Without the feature, the stock connect path runs unchanged.

Changed

• The speak feature now enables the websocket dependencies (tungstenite, tokio-tungstenite), and DeepgramError::WsError is available under speak as well as listen. No API change for default-feature users.

Fixed

• The Flux connection workers (STT and TTS) now end on the first terminal transport error (forwarding it exactly once) and complete the WebSocket closing handshake promptly when the server initiates the close, instead of leaving the close acknowledgement queued until socket teardown. After the session ends — peer close or terminal error — sends from the handle (FluxHandle::send_data/configure/force_end_turn, FluxSpeakHandle::speak/flush/interrupt/configure_speed) return an error instead of silently discarding the message.

Full changelog: 0.10.0...0.10.1