From da25fa2aa9a2629d7fe42e303c17e87279d65799 Mon Sep 17 00:00:00 2001
From: Daniel Howells <daniel@danielhowells.com>
Date: Sat, 13 Jun 2026 19:42:11 +0100
Subject: [PATCH] feat: add Miso TTS CLI

---
 generator.py |  4 +++-
 tts_cli.py   | 44 ++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 47 insertions(+), 1 deletion(-)
 create mode 100644 tts_cli.py

diff --git a/generator.py b/generator.py
index 8441cee..2084778 100644
--- a/generator.py
+++ b/generator.py
@@ -31,7 +31,9 @@ def load_llama3_tokenizer():
     """
     https://github.com/huggingface/transformers/issues/22794#issuecomment-2092623992
     """
-    tokenizer_name = "meta-llama/Llama-3.2-1B"
+    # meta-llama/Llama-3.2-1B is a gated repo; default to an ungated mirror with
+    # the identical tokenizer so no HF auth / license acceptance is needed.
+    tokenizer_name = os.environ.get("MISO_LLAMA_TOKENIZER", "unsloth/Llama-3.2-1B")
     tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
     bos = tokenizer.bos_token
     eos = tokenizer.eos_token
diff --git a/tts_cli.py b/tts_cli.py
new file mode 100644
index 0000000..99872b2
--- /dev/null
+++ b/tts_cli.py
@@ -0,0 +1,44 @@
+"""Isolated MisoTTS (8B) single-utterance generation CLI.
+
+Runs in its own venv (transformers 4.49). MisoTTS skips MPS (float64 limits),
+so it runs on CPU — slow for an 8B model, but functional. tilde's main server
+invokes this as a subprocess.
+
+Usage: python tts_cli.py --text "..." --out out.wav [--speaker 0]
+"""
+
+from __future__ import annotations
+
+import argparse
+import os
+import sys
+
+os.environ.setdefault("NO_TORCH_COMPILE", "1")
+os.environ.setdefault("HF_HUB_DOWNLOAD_TIMEOUT", "60")
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--text", required=True)
+    ap.add_argument("--out", required=True)
+    ap.add_argument("--speaker", type=int, default=0)
+    ap.add_argument("--max-ms", type=int, default=20000)
+    args = ap.parse_args()
+
+    import torch
+    import torchaudio
+    from generator import load_miso_8b
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    repo = os.environ.get("MISO_TTS_8B_MODEL", "MisoLabs/MisoTTS")
+    gen = load_miso_8b(device, model_path_or_repo_id=repo)
+    audio = gen.generate(
+        text=args.text, speaker=args.speaker, context=[], max_audio_length_ms=args.max_ms
+    )
+    torchaudio.save(args.out, audio.unsqueeze(0).cpu(), gen.sample_rate)
+    print(f"OK {args.out}", file=sys.stderr)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())