feat(aws): add auto language detection and mid-stream language switch…

[cldsime]

Summary

Adds support for Amazon Transcribe's automatic language identification parameters to the livekit-plugins-aws STT plugin, enabling auto language detection and mid-stream language switching without requiring users to manually specify a language_code.

Changes

6 new parameters added to STTOptions and STT.__init__():

• identify_language — detect the dominant language for the stream
• identify_multiple_languages — detect language switches mid-stream
• language_options — comma-separated list of expected language codes (2–12 required)
• preferred_language — bias detection toward a specific language
• vocabulary_names — custom vocabularies per language
• vocabulary_filter_names — vocabulary filters per language

All default to disabled (False / NOT_GIVEN), preserving full backward compatibility.

Conditional config building in SpeechStream._run():

language_code and identify_language/identify_multiple_languages are mutually exclusive per the AWS API. The config builder now conditionally sends one or the other:

if self._opts.identify_language:
    live_config["identify_language"] = True
    ...
elif self._opts.identify_multiple_languages:
    live_config["identify_multiple_languages"] = True
    ...
else:
    live_config["language_code"] = self._opts.language

Bug fix: filtered_config boolean handling

The original filter {k: v for k, v in live_config.items() if v and is_given(v)} silently drops False booleans. Replaced with explicit type checking that correctly preserves booleans, numbers, and NOT_GIVEN values.

Usage

# Existing behavior — unchanged
stt = STT(language="en-US")
Single language auto-detection
stt = STT(identify_language=True, language_options="en-US,es-US,fr-FR")
Multi-language mid-stream switching
stt = STT(

identify_multiple_languages=True,

language_options="en-US,es-US,fr-FR,de-DE,ja-JP,ko-KR,zh-HK,pt-BR,hi-IN,vi-VN,pl-PL,ru-RU",

)

Test Results

Tested with identify_multiple_languages=True and 12 language codes. All 12 configured languages were successfully detected across test sessions with mid-stream switching:

Language	Code	Detected
English	en-US	✅
Spanish	es-US	✅
French	fr-FR	✅
German	de-DE	✅
Japanese	ja-JP	✅
Korean	ko-KR	✅
Cantonese	zh-HK	✅
Portuguese	pt-BR	✅
Hindi	hi-IN	✅
Vietnamese	vi-VN	✅
Polish	pl-PL	✅
Russian	ru-RU	✅

Sample output showing mid-stream language switching in a single session:

[FINAL] [en-US] Good afternoon, everyone. Welcome to day 4.
[FINAL] [zh-HK] 你聽緊嘅係SBS電台廣東話節目
[FINAL] [vi-VN] Xin kính chào quý vị và các bạn
[FINAL] [pt-BR] Nós estamos aqui de azul hoje porque azul é a cor mais celebrando
[FINAL] [pl-PL] Dzień dobry, przy mikrofonie Joanna Borkowska Surucić

Backward Compatibility

All new parameters default to disabled. Existing code continues to work without any changes.

---

[CLAassistant]

All committers have signed the CLA.

[devin-ai-integration]

Devin Review found 1 potential issue.

View 4 additional findings in Devin Review.

[devin-ai-integration]

🔴 LanguageCode(None) crash when language identification is enabled and resp.language_code is missing

When identify_language or identify_multiple_languages is True, self._opts.language is set to None (line 142-144 in the constructor). In _streaming_recognize_response_to_speech_data, line 431 uses LanguageCode(resp.language_code or self._opts.language). If resp.language_code is also None or empty (which can happen for partial results), the or expression evaluates to None, and LanguageCode(None) is called. This crashes in _normalize_language (language.py:37) with AttributeError: 'NoneType' object has no attribute 'strip'.

(Refers to line 431)

---

Was this helpful? React with 👍 or 👎 to provide feedback.