pyenv virtualenv 3.14.5 speech
pyenv shell speech
pip install -r requirements.txt
# IBM Granite Speech (en, fr, de, es, pt, ja)
python ibm-granite/granite-speech/granite_speech_torch.py [audio_path]
# Cohere Transcribe (14 languages incl. vi) — text only, no timestamps
python cohere-labs/cohere-transcribe/cohere_transcribe_torch.py [audio_path] [language]
# Qwen3-ASR (30 languages incl. vi) — word timestamps via forced aligner, writes .srt
python qwen/qwen3-asr/qwen3_asr_torch.py [audio_path] [language]
# NVIDIA Nemotron Speech Streaming (en only) — segment/word timestamps, writes .srt
# requires: pip install "nemo_toolkit[asr]"
python nvidia/nemotron-speech-streaming/nemotron_streaming_torch.py [audio_path]
# Gemma 4 audio understanding (transcription/translation, 30s max per chunk — auto-chunked)
# via oMLX OpenAI-compatible server (default http://127.0.0.1:1234/v1, model gemma-4-12B-it-8bit)
# override with OMLX_HOST / OMLX_API_KEY / GEMMA_MODEL
python google/gemma-4/gemma4_audio_omlx.py [audio_path] [language]
# local transformers variant: default google/gemma-4-E4B-it (better Vietnamese than the oMLX path)
# override with GEMMA_MODEL=google/gemma-4-12B-it (needs ~24GB+ free RAM)
python google/gemma-4/gemma4_audio_torch.py [audio_path] [language]
# Higgs Audio v3 TTS (sglang-omni server, default http://localhost:8000) — voice-cloned Vietnamese speech
# serves the reference voice over a temp HTTP server so the TTS host can fetch it
# set HIGGS_HOST to point at a remote server, e.g. HIGGS_HOST=http://<server-ip>:8000
python higgs-audio/client/higgs_tts_vi.py [text_file] [ref_audio] [ref_text_file] [output_wav]
# VieNeu-TTS v3 Turbo (on-device Vietnamese TTS, 48 kHz, torch-free ONNX on CPU)
# runs locally — first call downloads the ~0.1B model from HF and caches it
# needs its own venv (see requirements.txt): pyenv virtualenv 3.14.5 vieneu && pip install vieneu
python pnnbao-ump/vieneu-tts/vieneu_tts_onnx.py [text_or_txt_file] [--voice NAME | --clone ref.wav] [-o out.wav]
python pnnbao-ump/vieneu-tts/vieneu_tts_onnx.py --list # list the 10 built-in preset voices
⚠️ Dependency conflict: cohere/gemma need transformers>=5.4, qwen-asr pins ==4.57.6,
nemo_toolkit wants 4.53.x — requirements.txt covers the >=5.4 family;
install qwen-asr and nemo_toolkit in their own virtualenvs (see comments in requirements.txt).
VieNeu-TTS is torch-free but pulls its own stack (onnxruntime, gradio, sea-g2p),
so install it in a separate vieneu virtualenv too.