[RaphaelFakhri]

The CLI accepts language names for --language (for example --language Japanese), and transcribe() accepts them too, because get_tokenizer() maps names to codes. The tokenizer and decoder then use the code, but transcribe() returns the original string, so result["language"] is "Japanese" instead of "ja", and "EN" stays "EN". The JSON output carries the same value. When the language is detected, or given as a lowercase code, the result holds a code, so anything that reads result["language"] sees two different formats.

Changes

• transcribe() now lowercases the requested language and maps a name to its code through TO_LANGUAGE_CODE before it builds the tokenizer, and stores that code in decode_options. A name such as "English" or "english" and a code in any case such as "EN" all produce "en". Unsupported values still raise the same ValueError from get_tokenizer().
• Adds test_language_is_reported_as_a_code, which transcribes jfk.flac with the tiny model using language="English" and language="EN", and checks that result["language"] == "en".

Reproduce

whisper tests/jfk.flac --model tiny --language English -f json -o out --fp16 False
python -c "import json; print(json.load(open('out/jfk.json'))['language'])"

Before: English. After: en.

Testing

• The new test fails on main (assert 'English' == 'en', assert 'EN' == 'en') and passes with this change.
• pytest -k 'not test_transcribe or test_transcribe[tiny] or test_transcribe[tiny.en]' -m 'not requires_cuda' passes.
• black, isort, and flake8 with the pre-commit settings report no issues.