Add support for Typhoon ASR Real-Time model with transcribe() function

[Copilot]

Adds scb10x/typhoon-asr-realtime model support to pythaiasr with comprehensive real-time inference capabilities.

Changes

• Model integration: Added Typhoon ASR to supported models list. Uses NVIDIA NeMo toolkit instead of HuggingFace transformers (existing models unaffected).

• Audio processing: Typhoon models require different handling:

  • Accepts file paths or numpy arrays
  • Resamples to 16kHz and normalizes audio via np.max(np.abs(data))
  • Uses temporary WAV files for NeMo's transcribe() API

• Real-time inference function: Added new transcribe() function specifically for Typhoon ASR models that provides:

  • Word-level timestamps (estimated, uniformly distributed)
  • Processing time metrics
  • Audio duration information
  • Rich structured output as dictionary instead of string
  • Device flexibility (CPU, CUDA, or auto)

• Dependencies: Added optional [typhoon] extra with nemo-toolkit[asr]>=1.21.0 and librosa

• Error handling: Raises informative ImportError when nemo-toolkit is missing, directing users to install via pip install pythaiasr[typhoon]

• Tests: Added unit tests for model listing, import error handling, and the new transcribe() function. Integration tests available via TEST_TYPHOON_ASR=true env var.

Usage

Basic ASR (string output)

from pythaiasr import asr

# File path or numpy array
result = asr("audio.wav", model="scb10x/typhoon-asr-realtime", device="cpu")

Real-time inference with detailed output

from pythaiasr import transcribe

# Basic transcription
result = transcribe("audio.wav", model="scb10x/typhoon-asr-realtime")
print(result['text'])
print(f"Processing time: {result['processing_time']:.2f}s")

# With word-level timestamps
result = transcribe("audio.wav", with_timestamps=True)
for ts in result['timestamps']:
    print(f"{ts['word']}: {ts['start']:.2f}s - {ts['end']:.2f}s")

Installation requires the typhoon extra:

pip install pythaiasr[typhoon]

Note: Timestamps are estimated by uniformly distributing words across the audio duration and may not reflect actual word timing in natural speech with pauses or variable rates.

Original prompt

---

This section details on the original issue you should resolve

<issue_title>Add support typhoon-asr-realtime</issue_title>
<issue_description>Typhoon ASR Real-Time is a next-generation, open-source Automatic Speech Recognition (ASR) model built specifically for real-world streaming applications in the Thai language. It is designed to deliver fast and accurate transcriptions while running efficiently on standard CPUs. This enables users to host their own ASR service, reducing costs and avoiding the need to send sensitive data to third-party cloud services. The model is trained on 10,000 hours of Thai audio transcriptions to help it generalize to any environments.

HuggingFace model: https://huggingface.co/typhoon-ai/typhoon-asr-realtime

GitHub: https://github.com/scb-10x/typhoon-asr</issue_description>

Comments on the Issue (you are @copilot in this section)

• Fixes Add support typhoon-asr-realtime #18

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.

[wannaphong]

@copilot Add ASR Real-Time Inference

[Copilot]

@copilot Add ASR Real-Time Inference

Added transcribe() function for real-time inference with Typhoon ASR. This provides word-level timestamps, processing time metrics, and audio duration information.

Example usage:

from pythaiasr import transcribe

result = transcribe("audio.wav", with_timestamps=True)
print(result['text'])
for ts in result['timestamps']:
    print(f"{ts['word']}: {ts['start']:.2f}s - {ts['end']:.2f}s")

Commit: 0ea8e24