feat(gladia, soniox): add translation support with input_language/input_text on SpeechData

[MSameerAbbas]

Summary

Adds real-time translation support to the Soniox and Gladia STT plugins by introducing input_language and input_text fields on the core SpeechData dataclass. Also includes a cleanup of the Soniox SpeechStream to align with patterns used by other plugins like Deepgram.

Closes #4943, closes #4402. Supersedes #5148.

Core type change

Two new optional fields on SpeechData in livekit-agents/livekit/agents/stt/stt.py:

• input_language: LanguageCode | None -- the detected/input language spoken by the user. Populated by STT services that support translation, where language holds the target language and input_language holds the original spoken language.
• input_text: str | None -- the original transcription in the input language, when translation is active.

Both default to None so existing behavior is completely unchanged.

Soniox: new features + SpeechStream cleanup

New features:

• Real-time translation (one-way and two-way) via TranslationConfig dataclass with __post_init__ validation
• Configurable max_endpoint_delay_ms (500-3000ms) for tuning endpoint detection latency
• models.py with Literal type aliases (SonioxRTModels, SonioxLanguages) for IDE autocomplete -- follows the same pattern as the Google STT plugin
• Flush sentinel mapped to Soniox's documented FINALIZE_MSG for clean session shutdown. Previously not handled.
• End-of-stream signal sent after input channel closes for graceful server-side shutdown

SpeechStream cleanup (while adding translation, simplified the streaming implementation to match other plugins):

• Consolidated into a single _run() that connects, runs tasks, and cleans up. The base class _main_task() already handles retry logic, so the plugin doesn't need its own retry loop.
• Reduced task count (4 -> 3): the intermediate audio_queue between _prepare_audio_task and _send_audio_task was consolidated into a single _send_task that reads _input_ch directly.
• Subtasks receive the WebSocket as a parameter rather than reading self._ws, similar to how the Deepgram plugin passes connection state.
• Server errors now raise APIConnectionError (5xx) or APIStatusError (4xx) so the base class can decide whether to retry. Unexpected WebSocket closure raises instead of silently returning.
• END_OF_SPEECH decoupled from FINAL_TRANSCRIPT in flush_endpoint -- previously both were gated on final text presence, so an error arriving mid-speech (after interim tokens but before finalization) would skip END_OF_SPEECH, leaving downstream consumers stuck in speaking state. Pre-existing bug also present on main. Only affects turn_detection="stt" (no VAD).

Translation design: Dict-keyed accumulators ("original" / "translation") route tokens by translation_status. At output time, _pick_primary selects the translation accumulator if it has content, otherwise falls back to original. _build_speech_data attaches input_text/input_language from the original accumulator when the primary is a translation. When translation is off, all tokens route to "original" and the identity check (primary is not original) skips the input fields -- one code path, no flags, no branching.

Gladia: translation fields

Surgical addition of input_language and input_text to the existing translation handler in _process_gladia_message. Extracts original utterance language and text from the translation message data and attaches them to the SpeechData. No structural changes.

What was NOT changed

• _TokenAccumulator class -- kept as-is, added a merge classmethod
• STT class -- kept as-is
• All STTOptions defaults preserved (model, sample_rate, num_channels, etc.)
• Context dataclasses (ContextObject, ContextGeneralItem, ContextTranslationTerm) -- unchanged
• Gladia plugin structure -- no cleanup, only the translation field addition

Files changed

File	Change
livekit-agents/livekit/agents/stt/stt.py	Add input_language, input_text to SpeechData
livekit-plugins/.../soniox/stt.py	SpeechStream rewrite + translation support
livekit-plugins/.../soniox/__init__.py	Export TranslationConfig, SonioxLanguages, SonioxRTModels
livekit-plugins/.../soniox/models.py	New file with Literal type aliases
livekit-plugins/.../gladia/stt.py	Add input_language/input_text to translation handler

Test plan

• Two-way translation (en/ur) -- verified both directions produce correct input_text and input_language
• One-way translation (to ur) -- verified single target language translation
• No translation (backward compat) -- verified input_text and input_language are None, identical to previous behavior
• max_endpoint_delay_ms -- verified API accepts the parameter
• TranslationConfig validation -- verified __post_init__ catches missing required fields
• END_OF_SPEECH lifecycle -- verified flush_endpoint emits END_OF_SPEECH independently of final text presence
• Ruff format and lint -- all checks passed
• mypy strict -- 0 new errors (1 pre-existing across all STT plugins)
• Unit test suite (294 passed, 2 skipped, 9 errors from missing LiveKit server -- pre-existing)

Refs: #4943, #4402

[MSameerAbbas]

Hey @tinalenguyen, I saw this was assigned to you - hope it's helpful! Would love your review.