zrb-extras 2.0.0

Collection of Zrb additional utilities

Zrb extras

zrb-extras is a pypi package.

You can install zrb-extras by invoking the following command:

pip install zrb-extras

Let your LLMTask speak and listen

Prerequisites

Termux

First of all, make sure termux has permission to access microphone/speaker

pkg update && pkg upgrade -y
pkg install pulseaudio termux-api -y

Run the following script or add it to ~/.bashrc

# start PulseAudio daemon
pulseaudio --start --load="module-native-protocol-tcp auth-ip-acl=127.0.0.1 auth-anonymous=1" --exit-idle-time=-1

# load module now (if it errors, check you gave Termux:API mic permission and restart Termux)
pactl load-module module-sles-source
# confirm source exists
pactl list short sources

# Start proot-distro
proot-distro login ubuntu

Proot-distro (Ubuntu)

apt install libasound2-dev portaudio19-dev pulseaudio

Create zrb_init.py

import os
from zrb.builtin import llm_ask
from zrb import llm_config
from zrb_extras.llm.tool import create_listen_tool, create_speak_tool

# Valid modes: "google", "openai", "termux", "vosk"
VOICE_MODE = os.getenv("VOICE_MODE", "vosk")
if VOICE_MODE not in ("google", "openai", "termux", "vosk"):
    VOICE_MODE = "vosk"

llm_ask.add_tool(
    create_speak_tool(
        mode=VOICE_MODE,
        genai_tts_model="gemini-2.5-flash-preview-tts",  # Optional
        genai_voice_name="Sulafat",  # Optional
        openai_tts_model="tts-1",  # Optional
        openai_voice_name="alloy",  # Optional
        sample_rate_out=24000,  # Optional
    )
)
llm_ask.add_tool(
    create_listen_tool(
        mode=VOICE_MODE,
        genai_stt_model="gemini-2.5-flash",  # Optional
        openai_stt_model="whisper-1",  # Optional
        sample_rate=16000,  # Optional
        channels=1,  # Optional
        silence_threshold=0.01,  # Optional
        max_silence=4.0,  # Optional
        # Sound Classification (optional)
        use_sound_classifier=True,  # Enable sound classification
        classification_model=None,  # Use default small model
        classification_system_prompt="Classify if the transcript contains actual speech or just background noise/fillers",
        classification_retries=2,  # Retry classification on failure
        fail_safe=True,  # Default to handling as speech if classification fails
    )
)

Sound Classification Feature

The create_listen_tool now includes an optional sound classification feature that uses an LLM to analyze transcripts and determine if they contain actual speech or just background noise, fillers, or non-speech sounds.

Key Features:

1. VAD is always used for initial speech detection (already implemented in existing listen tools)
2. When use_sound_classifier=True, transcripts are classified by an LLM using zrb's small model configuration system
3. Fail-safe default: If the classifier fails, it assumes the sound should be handled as speech
4. Structured output: Uses structured output types similar to ../zrb/src/zrb/task/llm/history_processor.py pattern
5. Configurable: Supports custom models, prompts, retries, and rate limiting

Usage Examples:

# Basic usage with sound classification
listen_tool = create_listen_tool(
    mode="vosk",
    use_sound_classifier=True,
    tool_name="smart_listen"
)

# With custom classification settings
listen_tool = create_listen_tool(
    mode="google",
    use_sound_classifier=True,
    classification_model="custom-model",
    classification_model_settings={"temperature": 0.1},
    classification_system_prompt="Classify speech vs noise",
    classification_retries=3,
    fail_safe=False,  # Raise exception on classification failure
    rate_limitter=my_rate_limiter,
    tool_name="custom_classifier_listen"
)

# Backward compatibility - old code still works
listen_tool = create_listen_tool(
    mode="termux",
    # No use_sound_classifier parameter
    tool_name="basic_listen"
)

How It Works:

1. The underlying listen tool (Vosk, Google, OpenAI, or Termux) captures audio and transcribes it
2. VAD (Voice Activity Detection) filters out silent periods
3. If use_sound_classifier=True, the transcript is sent to an LLM classifier
4. The classifier returns a structured response indicating:
   • is_speech: Boolean indicating if it's actual speech
   • confidence: Confidence score (0.0 to 1.0)
   • category: Optional category (e.g., "speech", "noise", "filler")
5. Based on the classification:
   • If is_speech=True: Returns the transcript
   • If is_speech=False: Returns empty string (ignores non-speech)

Benefits:

• Reduces false positives: Filters out background noise, coughs, throat clearing, etc.
• Improves accuracy: Only processes actual speech content
• Configurable: Can be tuned for different environments and use cases
• Backward compatible: Existing code continues to work without changes

Improving Voice Quality (Vosk Mode)

When using VOICE_MODE=vosk, speech recognition uses offline Vosk models and text-to-speech uses pyttsx3. Here's how to improve quality:

Vosk Speech Recognition Models

Recommended: For best accuracy, use the larger model. The default small model (~40MB) has limited accuracy.

Model	Size	Accuracy	Recommended
vosk-model-en-us-0.22	~1.8GB	Best	✅ Yes
vosk-model-en-us-daanzu-20200905	~1GB	Good	Good balance
vosk-model-small-en-us-0.15	~40MB	Limited	Default (not recommended)

Easiest way: Auto-download (recommended)

Vosk auto-downloads models to ~/.cache/vosk/ when you specify model_name. Just configure it in zrb_init.py:

listen = create_listen_tool(
    mode="vosk",
    vosk_model_name="vosk-model-en-us-0.22",  # Auto-downloads on first use
    # ... other options
)

Alternative: Manual download

If you prefer to pre-download (e.g., on a machine with better internet):

mkdir -p ~/.cache/vosk
cd ~/.cache/vosk
wget https://alphacephei.com/vosk/models/vosk-model-en-us-0.22.zip
unzip vosk-model-en-us-0.22.zip
rm vosk-model-en-us-0.22.zip

Alternative: Use vosk_model_path for custom locations:

listen_tool = create_listen_tool(
    mode="vosk",
    vosk_model_path="/custom/path/to/vosk-model-en-us-0.22",
)

pyttsx3 Text-to-Speech Quality

pyttsx3 uses your system's TTS engine. On Linux, it uses espeak/espeak-ng.

1. Install espeak-ng for better voices:

   # Ubuntu/Debian
   sudo apt install espeak-ng
   
   # Fedora
   sudo dnf install espeak-ng
   

2. List available voices:

   from zrb_extras.llm.tool.pyttsx3.speak import list_available_voices
   for voice in list_available_voices():
       print(f"{voice['id']}: {voice['name']}")
   

3. Configure voice via environment variables:

   # Set a specific voice (espeak-ng variants)
   export PYTTSX3_VOICE_NAME="english-us+m3"   # Male voice
   # export PYTTSX3_VOICE_NAME="english-us+f3" # Female voice
   
   # Adjust speed (words per minute, default 150)
   export PYTTSX3_VOICE_RATE="150"
   
   # Adjust volume (0.0 to 1.0, default 1.0)
   export PYTTSX3_VOICE_VOLUME="0.9"
   

4. Or pass to create_speak_tool:

   speak_tool = create_speak_tool(
       mode="vosk",
       voice_name="english-us+m3",  # Specific voice
       rate=150,                     # Words per minute
       volume=0.9,                   # Volume (0.0-1.0)
   )
   

macOS Users

On macOS, pyttsx3 falls back to the native say command which has better quality. You can use any installed macOS voice:

# List available voices
say -v ?

# Set voice
export PYTTSX3_VOICE_NAME="Samantha"  # Female voice
# export PYTTSX3_VOICE_NAME="Daniel"  # Male voice

# For maintainers

## Publish to pypi

To publish zrb-extras, you need to have a `Pypi` account:

- Log in or register to [https://pypi.org/](https://pypi.org/)
- Create an API token

You can also create a `TestPypi` account:

- Log in or register to [https://test.pypi.org/](https://test.pypi.org/)
- Create an API token

Once you have your API token, you need to configure poetry:

poetry config pypi-token.pypi

To publish zrb-extras, you can do the following command:

```bash
poetry publish --build

Updating version

You can update zrb-extras version by modifying the following section in pyproject.toml:

[project]
version = "0.0.2"

Adding dependencies

To add zrb-extras dependencies, you can edit the following section in pyproject.toml:

[project]
dependencies = [
    "Jinja2==3.1.2",
    "jsons==1.6.3"
]

Adding script

To make zrb-extras executable, you can edit the following section in pyproject.toml:

[project-scripts]
zrb-extras-hello = "zrb_extras.__main__:hello"

Now, whenever you run zrb-extras-hello, the main function on your __main__.py will be executed.