[hmellor]

What does this PR do?

Lets inference engines that run the processor themselves (e.g. vLLM's Transformers backend) use apply_transcription_request to build ASR prompts. This is a step towards serving /v1/audio/transcriptions through vLLM's Transformers backend instead of per-model native implementations such as vllm-project/vllm#62398 (VibeVoice-ASR).

• tokenize argument: added to apply_transcription_request in AudioFlamingo3, Canary, FunASR-Nano, GLM-ASR, MOSS-Transcribe-Diarize, Qwen3-ASR and VibeVoice-ASR. Defaults to True (unchanged). With False, the formatted prompts are returned as list[str] with unexpanded audio placeholders, to be processed later together with the audio. Voxtral already supports this.
• Truncated VibeVoice-ASR output: extract_speaker_dict raised JSONDecodeError when generation stopped mid-segment (e.g. at max_new_tokens). It now keeps the complete segments and logs a warning, in line with its "returning original text on failure" contract.

This overlaps with #49869 on the same signatures; whichever lands second needs a trivial rebase.

Tests: new VibeVoice-ASR processor tests check that processor(text=prompts, audio=audio) on the untokenized prompts matches the tokenized output, and that truncated output keeps its complete segments. Also checked against the microsoft/VibeVoice-ASR-HF processor: identical input_ids. The processor suites of all 7 models have the same results as main.

AI assistance (Claude Code) was used for this PR.

@eustlb @ebezzam

[hmellor]

With tokenize=False this returns list[str] rather than a BatchFeature, to match Voxtral's apply_transcription_request (and apply_chat_template, which ignores return_dict when not tokenizing). Happy to switch to a BatchFeature (e.g. {"text": ..., "audio": ...}, which would also keep audio loaded from URLs/paths) if you prefer.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: audioflamingo3, canary, fun_asr_nano, glmasr, moss_transcribe_diarize, qwen3_asr, vibevoice_asr

[eustlb]

thanks a lot! 🤗