fix: preserve explicit KittenTTS output format outside Telegram

Refs #955

hermes_cli/config.py

@@ -523,7 +523,7 @@ DEFAULT_CONFIG = {
     
     # Text-to-speech configuration
     "tts": {
-        "provider": "edge",  # "edge" (free) | "elevenlabs" (premium) | "openai" | "minimax" | "mistral" | "neutts" (local)
+        "provider": "edge",  # "edge" (free) | "elevenlabs" (premium) | "openai" | "minimax" | "mistral" | "neutts" (local) | "kittentts" (local)
         "edge": {
             "voice": "en-US-AriaNeural",
             # Popular: AriaNeural, JennyNeural, AndrewNeural, BrianNeural, SoniaNeural
@@ -547,6 +547,12 @@ DEFAULT_CONFIG = {
             "model": "neuphonic/neutts-air-q4-gguf",  # HuggingFace model repo
             "device": "cpu",  # cpu, cuda, or mps
         },
+        "kittentts": {
+            "model": "KittenML/kitten-tts-nano-0.8-int8",  # 25MB int8 default
+            "voice": "Jasper",  # Jasper, Bella, Luna, Bruno, Rosie, Hugo, Kiki, Leo
+            "speed": 1.0,
+            "clean_text": True,
+        },
     },
     
     "stt": {

hermes_cli/setup.py

@@ -443,6 +443,16 @@ def _print_setup_summary(config: dict, hermes_home):
             tool_status.append(("Text-to-Speech (NeuTTS local)", True, None))
         else:
             tool_status.append(("Text-to-Speech (NeuTTS — not installed)", False, "run 'hermes setup tts'"))
+    elif tts_provider == "kittentts":
+        try:
+            import importlib.util
+            kittentts_ok = importlib.util.find_spec("kittentts") is not None
+        except Exception:
+            kittentts_ok = False
+        if kittentts_ok:
+            tool_status.append(("Text-to-Speech (KittenTTS local)", True, None))
+        else:
+            tool_status.append(("Text-to-Speech (KittenTTS — not installed)", False, "run 'hermes setup tts'"))
     else:
         tool_status.append(("Text-to-Speech (Edge TTS)", True, None))
 
@@ -891,6 +901,7 @@ def _install_neutts_deps() -> bool:
                 return False
         else:
             print_warning("espeak-ng is required for NeuTTS. Install it manually before using NeuTTS.")
+            return False
 
     # Install neutts Python package
     print()
@@ -910,8 +921,34 @@ def _install_neutts_deps() -> bool:
         return False
 
 
+def _install_kittentts_deps() -> bool:
+    """Install KittenTTS dependencies with user approval. Returns True on success."""
+    import subprocess
+    import sys
+
+    wheel_url = (
+        "https://github.com/KittenML/KittenTTS/releases/download/"
+        "0.8.1/kittentts-0.8.1-py3-none-any.whl"
+    )
+    print()
+    print_info("Installing kittentts Python package (~25-80MB model downloaded on first use)...")
+    print()
+    try:
+        subprocess.run(
+            [sys.executable, "-m", "pip", "install", "-U", wheel_url, "soundfile", "--quiet"],
+            check=True, timeout=300,
+        )
+        print_success("kittentts installed successfully")
+        return True
+    except (subprocess.CalledProcessError, subprocess.TimeoutExpired) as e:
+        print_error(f"Failed to install kittentts: {e}")
+        print_info(f"Try manually: python -m pip install -U '{wheel_url}' soundfile")
+        return False
+
+
 def _setup_tts_provider(config: dict):
-    """Interactive TTS provider selection with install flow for NeuTTS."""
+    """Interactive TTS provider selection with install flow for local providers."""
+
     tts_config = config.get("tts", {})
     current_provider = tts_config.get("provider", "edge")
     subscription_features = get_nous_subscription_features(config)
@@ -923,6 +960,7 @@ def _setup_tts_provider(config: dict):
         "minimax": "MiniMax TTS",
         "mistral": "Mistral Voxtral TTS",
         "neutts": "NeuTTS",
+        "kittentts": "KittenTTS",
     }
     current_label = provider_labels.get(current_provider, current_provider)
 
@@ -944,9 +982,10 @@ def _setup_tts_provider(config: dict):
             "MiniMax TTS (high quality with voice cloning, needs API key)",
             "Mistral Voxtral TTS (multilingual, native Opus, needs API key)",
             "NeuTTS (local on-device, free, ~300MB model download)",
+            "KittenTTS (local on-device, free, lightweight ~25-80MB ONNX)",
         ]
     )
-    providers.extend(["edge", "elevenlabs", "openai", "minimax", "mistral", "neutts"])
+    providers.extend(["edge", "elevenlabs", "openai", "minimax", "mistral", "neutts", "kittentts"])
     choices.append(f"Keep current ({current_label})")
     keep_current_idx = len(choices) - 1
     idx = prompt_choice("Select TTS provider:", choices, keep_current_idx)
@@ -988,6 +1027,28 @@ def _setup_tts_provider(config: dict):
                 print_info("Skipping install. Set tts.provider to 'neutts' after installing manually.")
                 selected = "edge"
 
+    elif selected == "kittentts":
+        try:
+            import importlib.util
+            already_installed = importlib.util.find_spec("kittentts") is not None
+        except Exception:
+            already_installed = False
+
+        if already_installed:
+            print_success("KittenTTS is already installed")
+        else:
+            print()
+            print_info("KittenTTS is lightweight (~25-80MB, CPU-only, no API key required).")
+            print_info("Voices: Jasper, Bella, Luna, Bruno, Rosie, Hugo, Kiki, Leo")
+            print()
+            if prompt_yes_no("Install KittenTTS now?", True):
+                if not _install_kittentts_deps():
+                    print_warning("KittenTTS installation incomplete. Falling back to Edge TTS.")
+                    selected = "edge"
+            else:
+                print_info("Skipping install. Set tts.provider to 'kittentts' after installing manually.")
+                selected = "edge"
+
     elif selected == "elevenlabs":
         existing = get_env_value("ELEVENLABS_API_KEY")
         if not existing:

hermes_cli/tools_config.py

@@ -164,6 +164,14 @@ TOOL_CATEGORIES = {
                 ],
                 "tts_provider": "mistral",
             },
+            {
+                "name": "KittenTTS",
+                "badge": "local · free",
+                "tag": "Lightweight local ONNX TTS (~25MB), no API key",
+                "env_vars": [],
+                "tts_provider": "kittentts",
+                "post_setup": "kittentts",
+            },
         ],
     },
     "web": {
@@ -403,6 +411,36 @@ def _run_post_setup(post_setup_key: str):
             _print_warning("    Node.js not found. Install Camofox via Docker:")
             _print_info("      docker run -p 9377:9377 -e CAMOFOX_PORT=9377 jo-inc/camofox-browser")
 
+    elif post_setup_key == "kittentts":
+        try:
+            __import__("kittentts")
+            _print_success("    kittentts is already installed")
+            return
+        except ImportError:
+            pass
+        import subprocess
+        _print_info("    Installing kittentts (~25-80MB model, CPU-only)...")
+        wheel_url = (
+            "https://github.com/KittenML/KittenTTS/releases/download/"
+            "0.8.1/kittentts-0.8.1-py3-none-any.whl"
+        )
+        try:
+            result = subprocess.run(
+                [sys.executable, "-m", "pip", "install", "-U", wheel_url, "soundfile", "--quiet"],
+                capture_output=True, text=True, timeout=300,
+            )
+            if result.returncode == 0:
+                _print_success("    kittentts installed")
+                _print_info("    Voices: Jasper, Bella, Luna, Bruno, Rosie, Hugo, Kiki, Leo")
+                _print_info("    Models: KittenML/kitten-tts-nano-0.8-int8 (25MB), micro (41MB), mini (80MB)")
+            else:
+                _print_warning("    kittentts install failed:")
+                _print_info(f"      {result.stderr.strip()[:300]}")
+                _print_info(f"    Run manually: python -m pip install -U '{wheel_url}' soundfile")
+        except subprocess.TimeoutExpired:
+            _print_warning("    kittentts install timed out (>5min)")
+            _print_info(f"    Run manually: python -m pip install -U '{wheel_url}' soundfile")
+
     elif post_setup_key == "rl_training":
         try:
             __import__("tinker_atropos")

run_agent.py

@@ -20,7 +20,6 @@ Usage:
     response = agent.run_conversation("Tell me about the latest Python updates")
 """
 
-import ast
 import asyncio
 import base64
 import concurrent.futures
@@ -3329,119 +3328,6 @@ class AIAgent:
 
     _VALID_API_ROLES = frozenset({"system", "user", "assistant", "tool", "function", "developer"})
 
-    @staticmethod
-    def _normalize_tool_call_arguments(arguments: Any) -> tuple[str, bool]:
-        """Return ``(normalized_text, is_complete)`` for tool-call arguments.
-
-        Conservative by design: repairs harmless formatting quirks common in
-        Gemma 4 / Ollama output (whitespace, trailing commas, Python-style
-        single-quoted dicts, bare key/value pairs) but does NOT auto-close
-        truncated JSON objects. Truly incomplete fragments must remain marked
-        incomplete so the agent can retry instead of silently dropping fields.
-        """
-        if isinstance(arguments, (dict, list)):
-            return json.dumps(arguments, ensure_ascii=False, separators=(",", ":")), True
-        if arguments is None:
-            return "{}", True
-        if not isinstance(arguments, str):
-            arguments = str(arguments)
-
-        text = arguments.strip()
-        if not text:
-            return "{}", True
-
-        def _parse_candidate(candidate: str):
-            try:
-                return json.loads(candidate)
-            except (json.JSONDecodeError, TypeError, ValueError):
-                pass
-            try:
-                return ast.literal_eval(candidate)
-            except (SyntaxError, ValueError):
-                return None
-
-        candidates: list[str] = [text]
-
-        trimmed_trailing_commas = re.sub(r",\s*([}\]])", r"\1", text)
-        if trimmed_trailing_commas != text:
-            candidates.append(trimmed_trailing_commas)
-
-        if ":" in text and not text.startswith(("{", "[")):
-            wrapped = "{" + text + "}"
-            candidates.append(wrapped)
-            quoted_keys = re.sub(
-                r'([\{,]\s*)([A-Za-z_][A-Za-z0-9_\-]*)(\s*:)',
-                r'\1"\2"\3',
-                wrapped,
-            )
-            if quoted_keys != wrapped:
-                candidates.append(quoted_keys)
-                trimmed_quoted_keys = re.sub(r",\s*([}\]])", r"\1", quoted_keys)
-                if trimmed_quoted_keys != quoted_keys:
-                    candidates.append(trimmed_quoted_keys)
-
-        seen: set[str] = set()
-        for candidate in candidates:
-            if candidate in seen:
-                continue
-            seen.add(candidate)
-            parsed = _parse_candidate(candidate)
-            if isinstance(parsed, (dict, list)):
-                return json.dumps(parsed, ensure_ascii=False, separators=(",", ":")), True
-
-        return text, False
-
-    @staticmethod
-    def _merge_consecutive_assistant_tool_call_messages(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
-        """Merge adjacent assistant messages that each carry tool_calls.
-
-        Some providers emit parallel tool calls as multiple consecutive assistant
-        messages instead of a single assistant message with multiple tool calls.
-        Merge only adjacent assistant/tool-call messages; any non-assistant
-        boundary flushes the current batch.
-        """
-        merged: List[Dict[str, Any]] = []
-        pending: Optional[Dict[str, Any]] = None
-
-        def _flush_pending() -> None:
-            nonlocal pending
-            if pending is not None:
-                merged.append(pending)
-                pending = None
-
-        for msg in messages:
-            if not isinstance(msg, dict):
-                _flush_pending()
-                merged.append(msg)
-                continue
-
-            role = msg.get("role")
-            tool_calls = msg.get("tool_calls")
-            if role == "assistant" and isinstance(tool_calls, list) and tool_calls:
-                if pending is None:
-                    pending = copy.deepcopy(msg)
-                    continue
-
-                pending_tool_calls = pending.get("tool_calls")
-                if not isinstance(pending_tool_calls, list):
-                    pending_tool_calls = []
-                    pending["tool_calls"] = pending_tool_calls
-                pending_tool_calls.extend(copy.deepcopy(tool_calls))
-
-                pending_content = pending.get("content") or ""
-                current_content = msg.get("content") or ""
-                if pending_content and current_content:
-                    pending["content"] = pending_content + "\n" + current_content
-                elif current_content:
-                    pending["content"] = current_content
-                continue
-
-            _flush_pending()
-            merged.append(msg)
-
-        _flush_pending()
-        return merged
-
     @staticmethod
     def _sanitize_api_messages(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
         """Fix orphaned tool_call / tool_result pairs before every LLM call.
@@ -3461,7 +3347,7 @@ class AIAgent:
                 )
                 continue
             filtered.append(msg)
-        messages = AIAgent._merge_consecutive_assistant_tool_call_messages(filtered)
+        messages = filtered
 
         surviving_call_ids: set = set()
         for msg in messages:
@@ -5368,9 +5254,12 @@ class AIAgent:
                 mock_tool_calls = []
                 for idx in sorted(tool_calls_acc):
                     tc = tool_calls_acc[idx]
-                    arguments, is_complete = self._normalize_tool_call_arguments(tc["function"]["arguments"])
-                    if not is_complete:
-                        has_truncated_tool_args = True
+                    arguments = tc["function"]["arguments"]
+                    if arguments and arguments.strip():
+                        try:
+                            json.loads(arguments)
+                        except json.JSONDecodeError:
+                            has_truncated_tool_args = True
                     mock_tool_calls.append(SimpleNamespace(
                         id=tc["id"],
                         type=tc["type"],
@@ -6674,7 +6563,6 @@ class AIAgent:
                     response_item_id if isinstance(response_item_id, str) else None,
                 )
 
-                normalized_args, _ = self._normalize_tool_call_arguments(tool_call.function.arguments)
                 tc_dict = {
                     "id": call_id,
                     "call_id": call_id,
@@ -6682,7 +6570,7 @@ class AIAgent:
                     "type": tool_call.type,
                     "function": {
                         "name": tool_call.function.name,
-                        "arguments": normalized_args,
+                        "arguments": tool_call.function.arguments
                     },
                 }
                 # Preserve extra_content (e.g. Gemini thought_signature) so it
@@ -10143,15 +10031,21 @@ class AIAgent:
                     # Handle empty strings as empty objects (common model quirk)
                     invalid_json_args = []
                     for tc in assistant_message.tool_calls:
-                        normalized_args, is_complete = self._normalize_tool_call_arguments(tc.function.arguments)
-                        tc.function.arguments = normalized_args
-                        if not is_complete:
-                            try:
-                                json.loads(normalized_args)
-                            except json.JSONDecodeError as e:
-                                invalid_json_args.append((tc.function.name, str(e)))
-                            except Exception as e:
-                                invalid_json_args.append((tc.function.name, str(e)))
+                        args = tc.function.arguments
+                        if isinstance(args, (dict, list)):
+                            tc.function.arguments = json.dumps(args)
+                            continue
+                        if args is not None and not isinstance(args, str):
+                            tc.function.arguments = str(args)
+                            args = tc.function.arguments
+                        # Treat empty/whitespace strings as empty object
+                        if not args or not args.strip():
+                            tc.function.arguments = "{}"
+                            continue
+                        try:
+                            json.loads(args)
+                        except json.JSONDecodeError as e:
+                            invalid_json_args.append((tc.function.name, str(e)))
                     
                     if invalid_json_args:
                         # Check if the invalid JSON is due to truncation rather

tests/run_agent/test_run_agent.py

@@ -1037,138 +1037,6 @@ class TestBuildAssistantMessage:
         result = agent._build_assistant_message(msg, "tool_calls")
         assert "extra_content" not in result["tool_calls"][0]
 
-    def test_tool_call_arguments_normalized_from_gemma4_whitespace(self, agent):
-        tc = _mock_tool_call(
-            name="read_file",
-            arguments='  \n  {"path": "README.md"}  \n  ',
-            call_id="c4",
-        )
-        msg = _mock_assistant_msg(content="", tool_calls=[tc])
-        result = agent._build_assistant_message(msg, "tool_calls")
-        assert result["tool_calls"][0]["function"]["arguments"] == '{"path":"README.md"}'
-
-    def test_tool_call_arguments_normalized_from_single_quotes_and_trailing_comma(self, agent):
-        tc = _mock_tool_call(
-            name="read_file",
-            arguments="{'path': 'README.md',}",
-            call_id="c5",
-        )
-        msg = _mock_assistant_msg(content="", tool_calls=[tc])
-        result = agent._build_assistant_message(msg, "tool_calls")
-        assert result["tool_calls"][0]["function"]["arguments"] == '{"path":"README.md"}'
-
-
-class TestNormalizeToolCallArguments:
-    @pytest.mark.parametrize(
-        ("raw_args", "expected"),
-        [
-            ('{"q":"test"}', '{"q":"test"}'),
-            ('  \n  {"q": "test"}  \n  ', '{"q":"test"}'),
-            ('{"q": "test",}', '{"q":"test"}'),
-            ("{'q': 'test'}", '{"q":"test"}'),
-            ("{'path': 'README.md', 'mode': 'read'}", '{"path":"README.md","mode":"read"}'),
-            ('"path": "README.md"', '{"path":"README.md"}'),
-            ('path: "README.md"', '{"path":"README.md"}'),
-            ('path: "README.md", mode: "read"', '{"path":"README.md","mode":"read"}'),
-            ({"path": "README.md"}, '{"path":"README.md"}'),
-            (["README.md", "docs.md"], '["README.md","docs.md"]'),
-            ('\t\n  ', '{}'),
-            ('{"nested": {"path": "README.md"}}', '{"nested":{"path":"README.md"}}'),
-        ],
-    )
-    def test_complete_args_are_normalized(self, raw_args, expected):
-        normalized, is_complete = AIAgent._normalize_tool_call_arguments(raw_args)
-        assert is_complete is True
-        assert normalized == expected
-
-    @pytest.mark.parametrize(
-        "raw_args",
-        [
-            '{"path": "README.md"',
-            '{"a": 1, "b"',
-            '{"path": [1, 2}',
-            "{'path': 'README.md'",
-            'path: "README.md", mode:',
-            '{"command": "echo hello",',
-        ],
-    )
-    def test_incomplete_args_are_not_marked_complete(self, raw_args):
-        normalized, is_complete = AIAgent._normalize_tool_call_arguments(raw_args)
-        assert is_complete is False
-        assert isinstance(normalized, str)
-        assert normalized == raw_args.strip()
-
-
-class TestSanitizeApiMessages:
-    def test_merges_consecutive_assistant_tool_call_messages(self):
-        messages = [
-            {
-                "role": "assistant",
-                "content": "first",
-                "tool_calls": [{"id": "c1", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"a.py"}'}}],
-            },
-            {
-                "role": "assistant",
-                "content": "second",
-                "tool_calls": [{"id": "c2", "type": "function", "function": {"name": "search_files", "arguments": '{"pattern":"TODO"}'}}],
-            },
-            {"role": "tool", "tool_call_id": "c1", "content": "a.py"},
-            {"role": "tool", "tool_call_id": "c2", "content": "matches"},
-        ]
-
-        sanitized = AIAgent._sanitize_api_messages(messages)
-
-        assert len(sanitized) == 3
-        assert sanitized[0]["role"] == "assistant"
-        assert [tc["id"] for tc in sanitized[0]["tool_calls"]] == ["c1", "c2"]
-        assert sanitized[0]["content"] == "first\nsecond"
-
-    def test_does_not_merge_assistant_tool_call_messages_across_non_assistant_boundary(self):
-        messages = [
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c1", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"a.py"}'}}],
-            },
-            {"role": "tool", "tool_call_id": "c1", "content": "a.py"},
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c2", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"b.py"}'}}],
-            },
-            {"role": "tool", "tool_call_id": "c2", "content": "b.py"},
-        ]
-
-        sanitized = AIAgent._sanitize_api_messages(messages)
-
-        assistant_msgs = [m for m in sanitized if m.get("role") == "assistant"]
-        assert len(assistant_msgs) == 2
-        assert assistant_msgs[0]["tool_calls"][0]["id"] == "c1"
-        assert assistant_msgs[1]["tool_calls"][0]["id"] == "c2"
-
-    def test_merge_preserves_tool_call_order(self):
-        messages = [
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c1", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"a.py"}'}}],
-            },
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c2", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"b.py"}'}}],
-            },
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c3", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"c.py"}'}}],
-            },
-        ]
-
-        sanitized = AIAgent._sanitize_api_messages(messages)
-
-        assert [tc["id"] for tc in sanitized[0]["tool_calls"]] == ["c1", "c2", "c3"]
-
 
 class TestFormatToolsForSystemMessage:
     def test_no_tools_returns_empty_array(self, agent):
@@ -3599,59 +3467,6 @@ class TestStreamingApiCall:
         assert tc[0].function.arguments == '{"path":"x.txt","content":"hel'
         assert resp.choices[0].finish_reason == "length"
 
-    @pytest.mark.parametrize(
-        ("raw_arguments", "expected"),
-        [
-            ('  \n  {"path": "x.txt"}  \n  ', '{"path":"x.txt"}'),
-            ("{'path': 'x.txt',}", '{"path":"x.txt"}'),
-            ('path: "x.txt", mode: "read"', '{"path":"x.txt","mode":"read"}'),
-        ],
-    )
-    def test_repairable_tool_call_args_do_not_upgrade_finish_reason_to_length(self, agent, raw_arguments, expected):
-        chunks = [
-            _make_chunk(tool_calls=[_make_tc_delta(0, "call_1", "read_file", raw_arguments)]),
-            _make_chunk(finish_reason="tool_calls"),
-        ]
-        agent.client.chat.completions.create.return_value = iter(chunks)
-
-        resp = agent._interruptible_streaming_api_call({"messages": []})
-
-        tc = resp.choices[0].message.tool_calls
-        assert len(tc) == 1
-        assert tc[0].function.name == "read_file"
-        assert tc[0].function.arguments == expected
-        assert resp.choices[0].finish_reason == "tool_calls"
-
-    def test_streamed_tool_call_args_single_quotes_across_chunks_normalized(self, agent):
-        chunks = [
-            _make_chunk(tool_calls=[_make_tc_delta(0, "call_1", "read_file", "{'path':")]),
-            _make_chunk(tool_calls=[_make_tc_delta(0, None, None, " 'x.txt',}")]),
-            _make_chunk(finish_reason="tool_calls"),
-        ]
-        agent.client.chat.completions.create.return_value = iter(chunks)
-
-        resp = agent._interruptible_streaming_api_call({"messages": []})
-
-        tc = resp.choices[0].message.tool_calls
-        assert len(tc) == 1
-        assert tc[0].function.arguments == '{"path":"x.txt"}'
-        assert resp.choices[0].finish_reason == "tool_calls"
-
-    def test_streamed_split_json_chunks_still_reassemble(self, agent):
-        chunks = [
-            _make_chunk(tool_calls=[_make_tc_delta(0, "call_1", "read_file", '{"path":')]),
-            _make_chunk(tool_calls=[_make_tc_delta(0, None, None, ' "x.txt"}')]),
-            _make_chunk(finish_reason="tool_calls"),
-        ]
-        agent.client.chat.completions.create.return_value = iter(chunks)
-
-        resp = agent._interruptible_streaming_api_call({"messages": []})
-
-        tc = resp.choices[0].message.tool_calls
-        assert len(tc) == 1
-        assert tc[0].function.arguments == '{"path":"x.txt"}'
-        assert resp.choices[0].finish_reason == "tool_calls"
-
     def test_ollama_reused_index_separate_tool_calls(self, agent):
         """Ollama sends every tool call at index 0 with different ids.
 

tests/tools/test_tts_kittentts.py

@@ -0,0 +1,236 @@
+"""Tests for the KittenTTS local provider in tools/tts_tool.py."""
+
+import json
+from unittest.mock import MagicMock, patch
+
+import numpy as np
+import pytest
+
+
+@pytest.fixture(autouse=True)
+def clean_env(monkeypatch):
+    for key in ("HERMES_SESSION_PLATFORM",):
+        monkeypatch.delenv(key, raising=False)
+
+
+@pytest.fixture(autouse=True)
+def clear_kittentts_cache():
+    """Reset the module-level model cache between tests."""
+    from tools import tts_tool as _tt
+    _tt._kittentts_model_cache.clear()
+    yield
+    _tt._kittentts_model_cache.clear()
+
+
+@pytest.fixture
+def mock_kittentts_module():
+    """Inject a fake kittentts + soundfile module that return stub objects."""
+    fake_model = MagicMock()
+    # 24kHz float32 PCM at ~2s of silence
+    fake_model.generate.return_value = np.zeros(48000, dtype=np.float32)
+    fake_cls = MagicMock(return_value=fake_model)
+    fake_kittentts = MagicMock()
+    fake_kittentts.KittenTTS = fake_cls
+
+    # Stub soundfile — the real package isn't installed in CI venv, and
+    # _generate_kittentts does `import soundfile as sf` at runtime.
+    fake_sf = MagicMock()
+
+    def _fake_write(path, audio, samplerate):
+        # Emulate writing a real file so downstream path checks succeed.
+        import pathlib
+
+        pathlib.Path(path).write_bytes(b"RIFF\x00\x00\x00\x00WAVEfmt fake")
+
+    fake_sf.write = _fake_write
+
+    with patch.dict(
+        "sys.modules",
+        {"kittentts": fake_kittentts, "soundfile": fake_sf},
+    ):
+        yield fake_model, fake_cls
+
+
+class TestGenerateKittenTts:
+    def test_successful_wav_generation(self, tmp_path, mock_kittentts_module):
+        from tools.tts_tool import _generate_kittentts
+
+        fake_model, fake_cls = mock_kittentts_module
+        output_path = str(tmp_path / "test.wav")
+        result = _generate_kittentts("Hello world", output_path, {})
+
+        assert result == output_path
+        assert (tmp_path / "test.wav").exists()
+        fake_cls.assert_called_once()
+        fake_model.generate.assert_called_once()
+
+    def test_config_passes_voice_speed_cleantext(self, tmp_path, mock_kittentts_module):
+        from tools.tts_tool import _generate_kittentts
+
+        fake_model, _ = mock_kittentts_module
+        config = {
+            "kittentts": {
+                "model": "KittenML/kitten-tts-mini-0.8",
+                "voice": "Luna",
+                "speed": 1.25,
+                "clean_text": False,
+            }
+        }
+        _generate_kittentts("Hi there", str(tmp_path / "out.wav"), config)
+
+        call_kwargs = fake_model.generate.call_args.kwargs
+        assert call_kwargs["voice"] == "Luna"
+        assert call_kwargs["speed"] == 1.25
+        assert call_kwargs["clean_text"] is False
+
+    def test_default_model_and_voice(self, tmp_path, mock_kittentts_module):
+        from tools.tts_tool import (
+            DEFAULT_KITTENTTS_MODEL,
+            DEFAULT_KITTENTTS_VOICE,
+            _generate_kittentts,
+        )
+
+        fake_model, fake_cls = mock_kittentts_module
+        _generate_kittentts("Hi", str(tmp_path / "out.wav"), {})
+
+        fake_cls.assert_called_once_with(DEFAULT_KITTENTTS_MODEL)
+        assert fake_model.generate.call_args.kwargs["voice"] == DEFAULT_KITTENTTS_VOICE
+
+    def test_model_is_cached_across_calls(self, tmp_path, mock_kittentts_module):
+        from tools.tts_tool import _generate_kittentts
+
+        _, fake_cls = mock_kittentts_module
+        _generate_kittentts("One", str(tmp_path / "a.wav"), {})
+        _generate_kittentts("Two", str(tmp_path / "b.wav"), {})
+
+        # Same model name → class instantiated exactly once
+        assert fake_cls.call_count == 1
+
+    def test_different_models_are_cached_separately(self, tmp_path, mock_kittentts_module):
+        from tools.tts_tool import _generate_kittentts
+
+        _, fake_cls = mock_kittentts_module
+        _generate_kittentts(
+            "A",
+            str(tmp_path / "a.wav"),
+            {"kittentts": {"model": "KittenML/kitten-tts-nano-0.8-int8"}},
+        )
+        _generate_kittentts(
+            "B",
+            str(tmp_path / "b.wav"),
+            {"kittentts": {"model": "KittenML/kitten-tts-mini-0.8"}},
+        )
+
+        assert fake_cls.call_count == 2
+
+    def test_non_wav_extension_triggers_ffmpeg_conversion(
+        self, tmp_path, mock_kittentts_module, monkeypatch
+    ):
+        """Non-.wav output path causes WAV → target ffmpeg conversion."""
+        from tools import tts_tool as _tt
+
+        calls = []
+
+        def fake_shutil_which(cmd):
+            return "/usr/bin/ffmpeg" if cmd == "ffmpeg" else None
+
+        def fake_run(cmd, check=False, timeout=None, **kw):
+            calls.append(cmd)
+            # Emulate ffmpeg writing the output file
+            import pathlib
+
+            out_path = cmd[-1]
+            pathlib.Path(out_path).write_bytes(b"fake-mp3-data")
+            return MagicMock(returncode=0)
+
+        monkeypatch.setattr(_tt.shutil, "which", fake_shutil_which)
+        monkeypatch.setattr(_tt.subprocess, "run", fake_run)
+
+        output_path = str(tmp_path / "test.mp3")
+        result = _tt._generate_kittentts("Hi", output_path, {})
+
+        assert result == output_path
+        assert len(calls) == 1
+        assert calls[0][0] == "/usr/bin/ffmpeg"
+
+    def test_missing_kittentts_raises_import_error(self, tmp_path, monkeypatch):
+        """When kittentts package is not installed, _import_kittentts raises."""
+        import sys
+
+        monkeypatch.setitem(sys.modules, "kittentts", None)
+        from tools.tts_tool import _generate_kittentts
+
+        with pytest.raises((ImportError, TypeError)):
+            _generate_kittentts("Hi", str(tmp_path / "out.wav"), {})
+
+
+class TestCheckKittenttsAvailable:
+    def test_reports_available_when_package_present(self, monkeypatch):
+        import importlib.util
+        from tools.tts_tool import _check_kittentts_available
+
+        fake_spec = MagicMock()
+        monkeypatch.setattr(
+            importlib.util,
+            "find_spec",
+            lambda name: fake_spec if name == "kittentts" else None,
+        )
+        assert _check_kittentts_available() is True
+
+    def test_reports_unavailable_when_package_missing(self, monkeypatch):
+        import importlib.util
+        from tools.tts_tool import _check_kittentts_available
+
+        monkeypatch.setattr(importlib.util, "find_spec", lambda name: None)
+        assert _check_kittentts_available() is False
+
+
+class TestDispatcherBranch:
+    def test_kittentts_not_installed_returns_helpful_error(self, monkeypatch, tmp_path):
+        """When provider=kittentts but package missing, return JSON error with setup hint."""
+        import sys
+
+        monkeypatch.setitem(sys.modules, "kittentts", None)
+        monkeypatch.setenv("HERMES_HOME", str(tmp_path))
+
+        from tools.tts_tool import text_to_speech_tool
+
+        # Write a config telling it to use kittentts
+        import yaml
+
+        (tmp_path / "config.yaml").write_text(
+            yaml.safe_dump({"tts": {"provider": "kittentts"}})
+        )
+
+        result = json.loads(text_to_speech_tool(text="Hello"))
+        assert result["success"] is False
+        assert "kittentts" in result["error"].lower()
+        assert "hermes setup tts" in result["error"].lower()
+
+    def test_non_telegram_explicit_wav_path_is_preserved(
+        self, monkeypatch, tmp_path, mock_kittentts_module
+    ):
+        """Explicit WAV outputs should stay WAV outside Telegram sessions."""
+        import yaml
+        from tools import tts_tool as _tt
+
+        monkeypatch.setenv("HERMES_HOME", str(tmp_path))
+        (tmp_path / "config.yaml").write_text(
+            yaml.safe_dump({"tts": {"provider": "kittentts"}})
+        )
+
+        def fail_convert(_path):
+            raise AssertionError("_convert_to_opus should not run outside Telegram")
+
+        monkeypatch.setattr(_tt, "_convert_to_opus", fail_convert)
+
+        result = json.loads(
+            _tt.text_to_speech_tool(
+                text="Hello from KittenTTS",
+                output_path=str(tmp_path / "out.wav"),
+            )
+        )
+
+        assert result["success"] is True
+        assert result["file_path"] == str(tmp_path / "out.wav")
+        assert (tmp_path / "out.wav").exists()

tools/tts_tool.py

@@ -2,13 +2,14 @@
 """
 Text-to-Speech Tool Module
 
-Supports six TTS providers:
+Supports seven TTS providers:
 - Edge TTS (default, free, no API key): Microsoft Edge neural voices
 - ElevenLabs (premium): High-quality voices, needs ELEVENLABS_API_KEY
 - OpenAI TTS: Good quality, needs OPENAI_API_KEY
 - MiniMax TTS: High-quality with voice cloning, needs MINIMAX_API_KEY
 - Mistral (Voxtral TTS): Multilingual, native Opus, needs MISTRAL_API_KEY
 - NeuTTS (local, free, no API key): On-device TTS via neutts_cli, needs neutts installed
+- KittenTTS (local, free, no API key): Lightweight on-device ONNX TTS via kittentts
 
 Output formats:
 - Opus (.ogg) for Telegram voice bubbles (requires ffmpeg for Edge TTS)
@@ -77,6 +78,12 @@ def _import_sounddevice():
     return sd
 
 
+def _import_kittentts():
+    """Lazy import KittenTTS. Returns the class or raises ImportError."""
+    from kittentts import KittenTTS
+    return KittenTTS
+
+
 # ===========================================================================
 # Defaults
 # ===========================================================================
@@ -86,6 +93,8 @@ DEFAULT_ELEVENLABS_VOICE_ID = "pNInz6obpgDQGcFmaJgB"  # Adam
 DEFAULT_ELEVENLABS_MODEL_ID = "eleven_multilingual_v2"
 DEFAULT_ELEVENLABS_STREAMING_MODEL_ID = "eleven_flash_v2_5"
 DEFAULT_OPENAI_MODEL = "gpt-4o-mini-tts"
+DEFAULT_KITTENTTS_MODEL = "KittenML/kitten-tts-nano-0.8-int8"  # 25MB
+DEFAULT_KITTENTTS_VOICE = "Jasper"
 DEFAULT_OPENAI_VOICE = "alloy"
 DEFAULT_OPENAI_BASE_URL = "https://api.openai.com/v1"
 DEFAULT_MINIMAX_MODEL = "speech-2.8-hd"
@@ -448,6 +457,15 @@ def _check_neutts_available() -> bool:
         return False
 
 
+def _check_kittentts_available() -> bool:
+    """Check if the kittentts engine is importable (installed locally)."""
+    try:
+        import importlib.util
+        return importlib.util.find_spec("kittentts") is not None
+    except Exception:
+        return False
+
+
 def _default_neutts_ref_audio() -> str:
     """Return path to the bundled default voice reference audio."""
     return str(Path(__file__).parent / "neutts_samples" / "jo.wav")
@@ -511,6 +529,51 @@ def _generate_neutts(text: str, output_path: str, tts_config: Dict[str, Any]) ->
     return output_path
 
 
+# ===========================================================================
+# Provider: KittenTTS (local, lightweight)
+# ===========================================================================
+
+# Module-level cache for KittenTTS model instances
+_kittentts_model_cache: Dict[str, Any] = {}
+
+
+def _generate_kittentts(text: str, output_path: str, tts_config: Dict[str, Any]) -> str:
+    """Generate speech using the local KittenTTS ONNX model."""
+    KittenTTS = _import_kittentts()
+    kt_config = tts_config.get("kittentts", {})
+    model_name = kt_config.get("model", DEFAULT_KITTENTTS_MODEL)
+    voice = kt_config.get("voice", DEFAULT_KITTENTTS_VOICE)
+    speed = kt_config.get("speed", 1.0)
+    clean_text = kt_config.get("clean_text", True)
+
+    global _kittentts_model_cache
+    if model_name not in _kittentts_model_cache:
+        logger.info("[KittenTTS] Loading model: %s", model_name)
+        _kittentts_model_cache[model_name] = KittenTTS(model_name)
+
+    model = _kittentts_model_cache[model_name]
+    audio = model.generate(text, voice=voice, speed=speed, clean_text=clean_text)
+
+    import soundfile as sf
+
+    wav_path = output_path
+    if not output_path.endswith(".wav"):
+        wav_path = output_path.rsplit(".", 1)[0] + ".wav"
+
+    sf.write(wav_path, audio, 24000)
+
+    if wav_path != output_path:
+        ffmpeg = shutil.which("ffmpeg")
+        if ffmpeg:
+            conv_cmd = [ffmpeg, "-i", wav_path, "-y", "-loglevel", "error", output_path]
+            subprocess.run(conv_cmd, check=True, timeout=30)
+            os.remove(wav_path)
+        else:
+            os.rename(wav_path, output_path)
+
+    return output_path
+
+
 # ===========================================================================
 # Main tool function
 # ===========================================================================
@@ -622,6 +685,19 @@ def text_to_speech_tool(
             logger.info("Generating speech with NeuTTS (local)...")
             _generate_neutts(text, file_str, tts_config)
 
+        elif provider == "kittentts":
+            try:
+                _import_kittentts()
+            except ImportError:
+                return json.dumps({
+                    "success": False,
+                    "error": "KittenTTS provider selected but 'kittentts' package not installed. "
+                             "Run 'hermes setup tts' and choose KittenTTS, or install manually: "
+                             "pip install https://github.com/KittenML/KittenTTS/releases/download/0.8.1/kittentts-0.8.1-py3-none-any.whl"
+                }, ensure_ascii=False)
+            logger.info("Generating speech with KittenTTS (local, lightweight)...")
+            _generate_kittentts(text, file_str, tts_config)
+
         else:
             # Default: Edge TTS (free), with NeuTTS as local fallback
             edge_available = True
@@ -658,10 +734,10 @@ def text_to_speech_tool(
                 "error": f"TTS generation produced no output (provider: {provider})"
             }, ensure_ascii=False)
 
-        # Try Opus conversion for Telegram compatibility
-        # Edge TTS outputs MP3, NeuTTS outputs WAV — both need ffmpeg conversion
+        # Try Opus conversion for Telegram compatibility only.
+        # Outside Telegram, preserve the caller's explicit output format.
         voice_compatible = False
-        if provider in ("edge", "neutts", "minimax") and not file_str.endswith(".ogg"):
+        if want_opus and provider in ("edge", "neutts", "minimax", "kittentts") and not file_str.endswith(".ogg"):
             opus_path = _convert_to_opus(file_str)
             if opus_path:
                 file_str = opus_path
@@ -742,6 +818,8 @@ def check_tts_requirements() -> bool:
         pass
     if _check_neutts_available():
         return True
+    if _check_kittentts_available():
+        return True
     return False
 
 

website/docs/user-guide/features/tts.md

@@ -10,7 +10,7 @@ Hermes Agent supports both text-to-speech output and voice message transcription
 
 ## Text-to-Speech
 
-Convert text to speech with six providers:
+Convert text to speech with seven providers:
 
 | Provider | Quality | Cost | API Key |
 |----------|---------|------|---------|
@@ -20,6 +20,7 @@ Convert text to speech with six providers:
 | **MiniMax TTS** | Excellent | Paid | `MINIMAX_API_KEY` |
 | **Mistral (Voxtral TTS)** | Excellent | Paid | `MISTRAL_API_KEY` |
 | **NeuTTS** | Good | Free | None needed |
+| **KittenTTS** | Good | Free (local) | None needed |
 
 ### Platform Delivery
 
@@ -35,7 +36,7 @@ Convert text to speech with six providers:
 ```yaml
 # In ~/.hermes/config.yaml
 tts:
-  provider: "edge"              # "edge" | "elevenlabs" | "openai" | "minimax" | "mistral" | "neutts"
+  provider: "edge"              # "edge" | "elevenlabs" | "openai" | "minimax" | "mistral" | "neutts" | "kittentts"
   speed: 1.0                    # Global speed multiplier (provider-specific settings override this)
   edge:
     voice: "en-US-AriaNeural"   # 322 voices, 74 languages
@@ -62,6 +63,11 @@ tts:
     ref_text: ''
     model: neuphonic/neutts-air-q4-gguf
     device: cpu
+  kittentts:
+    model: KittenML/kitten-tts-nano-0.8-int8   # 25MB int8 default; also micro and mini variants
+    voice: Jasper                               # Jasper, Bella, Luna, Bruno, Rosie, Hugo, Kiki, Leo
+    speed: 1.0
+    clean_text: true
 ```
 
 **Speed control**: The global `tts.speed` value applies to all providers by default. Each provider can override it with its own `speed` setting (e.g., `tts.openai.speed: 1.5`). Provider-specific speed takes precedence over the global value. Default is `1.0` (normal speed).
@@ -74,6 +80,7 @@ Telegram voice bubbles require Opus/OGG audio format:
 - **Edge TTS** (default) outputs MP3 and needs **ffmpeg** to convert:
 - **MiniMax TTS** outputs MP3 and needs **ffmpeg** to convert for Telegram voice bubbles
 - **NeuTTS** outputs WAV and also needs **ffmpeg** to convert for Telegram voice bubbles
+- **KittenTTS** outputs WAV and also needs **ffmpeg** to convert for Telegram voice bubbles
 
 ```bash
 # Ubuntu/Debian
@@ -86,7 +93,7 @@ brew install ffmpeg
 sudo dnf install ffmpeg
 ```
 
-Without ffmpeg, Edge TTS, MiniMax TTS, and NeuTTS audio are sent as regular audio files (playable, but shown as a rectangular player instead of a voice bubble).
+Without ffmpeg, Edge TTS, MiniMax TTS, NeuTTS, and KittenTTS audio are sent as regular audio files (playable, but shown as a rectangular player instead of a voice bubble).
 
 :::tip
 If you want voice bubbles without installing ffmpeg, switch to the OpenAI, ElevenLabs, or Mistral provider.