Speech To Speech: Build voice agents with open-source models

A low-latency, fully modular voice-agent pipeline: VAD -> STT -> LLM -> TTS, exposed through an OpenAI Realtime-compatible WebSocket API. Every component is swappable. The LLM slot speaks OpenAI-compatible protocols, so you can point it at a hosted provider, at HF Inference Providers, or at a vLLM or llama.cpp server on your own hardware for a fully local, fully open stack.

This pipeline runs in production as the conversation backend for thousands of Reachy Mini robots.

pip install speech-to-speech
export OPENAI_API_KEY=...
speech-to-speech serve

This starts an OpenAI Realtime-compatible server at ws://localhost:8765/v1/realtime using Parakeet TDT for local STT, an OpenAI-compatible LLM, and Qwen3-TTS for local speech output.

Talk to it from a second terminal:

speech-to-speech talk --url ws://127.0.0.1:8765/v1/realtime

To start the server and packaged microphone/speaker client in one command:

speech-to-speech local

Prefer to keep the LLM on your own machine? Serve Gemma 4 with llama.cpp:

llama-server -hf ggml-org/gemma-4-E4B-it-GGUF -np 2 -c 65536 -fa on --swa-full

Then point the OpenAI-compatible LLM backend at it:

speech-to-speech serve \
    --model_name "ggml-org/gemma-4-E4B-it-GGUF" \
    --responses_api_base_url "http://127.0.0.1:8080/v1" \
    --responses_api_api_key ""

Any OpenAI Realtime-compatible client can connect. See Realtime API for the protocol and LLM backends for provider and local-server options.

• How it works
• Installation
• Supported components
• Commands
• Realtime API
• LLM backends
• Multi-language support
• Pocket TTS
• CLI reference
• Contributing
• Star history
• Citations

The pipeline is a cascade of four components, each running in its own thread and connected by queues:

1. Voice Activity Detection (VAD): Silero VAD v5 detects speech boundaries and turn-taking.
2. Speech to Text (STT): transcribes the user's turn, with optional live partial transcripts.
3. Language Model (LLM): generates the response, streaming text and tool calls.
4. Text to Speech (TTS): synthesizes audio and streams it back to the client.

Every stage has multiple interchangeable backends, selected via CLI flags. The code is designed for easy modification, with a focus on models available through Transformers and the Hugging Face Hub.

Requires Python 3.10+.

pip install speech-to-speech

The default install covers the standard realtime path:

• Parakeet TDT for STT
• OpenAI-compatible API for the language model
• Qwen3-TTS for speech output, using the GGML backend by default on non-macOS platforms and mlx-audio on Apple Silicon
• local audio and realtime server modes

macOS and non-macOS dependencies are resolved automatically via platform markers in pyproject.toml.

On Linux, the Qwen3-TTS GGML backend comes from faster-qwen3-tts[ggml]. Its default qwentts-cpp-python wheel on PyPI targets CUDA 12.8. If your machine does not have the CUDA 12 runtime that wheel expects, install the matching wheel from the Hugging Face wheelhouse before installing speech-to-speech:

# CUDA 13.x
pip install "qwentts-cpp-python==0.3.1+cu130" \
  -f https://huggingface.co/datasets/andito/qwentts-cpp-python-wheels/tree/main/whl/cu130

# CUDA 12.4
pip install "qwentts-cpp-python==0.3.1+cu124" \
  -f https://huggingface.co/datasets/andito/qwentts-cpp-python-wheels/tree/main/whl/cu124

# CPU-only fallback
pip install "qwentts-cpp-python==0.3.1+cpu" \
  -f https://huggingface.co/datasets/andito/qwentts-cpp-python-wheels/tree/main/whl/cpu

pip install speech-to-speech

To use the previous CUDA-graphs implementation instead of GGML, pass --qwen3_tts_backend torch.

Optional components are installed with pip extras:

pip install "speech-to-speech[kokoro]"          # Kokoro-82M TTS on non-macOS
pip install "speech-to-speech[pocket]"          # Pocket TTS
pip install "speech-to-speech[chattts]"         # ChatTTS
pip install "speech-to-speech[faster-whisper]"  # Faster Whisper STT
pip install "speech-to-speech[whisper-mlx]"     # Lightning Whisper MLX STT on macOS
pip install "speech-to-speech[paraformer]"      # Paraformer STT through FunASR
pip install "speech-to-speech[mlx-lm]"          # mlx-vlm support for vision models on macOS

Deprecated implementations, including MeloTTS, live in archive/ and are no longer wired into the CLI.

Note on DeepFilterNet: DeepFilterNet, used for optional audio enhancement in VAD, requires numpy<2 and conflicts with Pocket TTS, which requires numpy>=2. Install it manually only in environments where you are not using Pocket TTS.

git clone https://github.com/huggingface/speech-to-speech.git
cd speech-to-speech
uv sync

This installs the package in editable mode and makes the speech-to-speech CLI available.

Select implementations with --stt, --llm_backend, and --tts. Run speech-to-speech serve -h for exact values and backend-specific flags.

serve binds to 127.0.0.1 by default; pass --host 0.0.0.0 explicitly for network exposure. local always binds to loopback and connects the same packaged client at ws://127.0.0.1:<port>/v1/realtime.

--mode is deprecated and will stop working soon. During this migration window, speech-to-speech --mode realtime runs speech-to-speech serve, and speech-to-speech --mode local runs speech-to-speech local; both print a warning. All other mode values have been removed and exit with guidance to use the new commands.

export OPENAI_API_KEY=...
speech-to-speech serve

This is equivalent to:

speech-to-speech serve \
    --thresh 0.6 \
    --stt parakeet-tdt \
    --llm_backend responses-api \
    --tts qwen3 \
    --qwen3_tts_model_name Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice \
    --qwen3_tts_speaker Aiden \
    --qwen3_tts_language auto \
    --qwen3_tts_backend ggml \
    --qwen3_tts_non_streaming_mode True \
    --qwen3_tts_mlx_quantization 6bit \
    --model_name gpt-5.4-mini \
    --chat_size 30 \
    --responses_api_stream \
    --enable_live_transcription

The default model is gpt-5.4-mini through the OpenAI Responses API. Override it with --model_name, and set --responses_api_base_url for another OpenAI-compatible provider or server.

speech-to-speech local --mac-optimal-settings

Optionally with a specific LLM:

speech-to-speech local \
    --mac-optimal-settings \
    --model_name mlx-community/Qwen3-4B-Instruct-2507-bf16

This setting:

• Uses MPS defaults for supported model components.
• Sets Parakeet TDT for STT.
• Sets MLX LM as the LLM backend.
• Sets Qwen3-TTS for TTS, using mlx-audio with the 6bit MLX variant by default.

The preset supplies these as defaults only: explicit --device, component-device flags such as --qwen3_tts_device, and --stt, --llm_backend, --model_name, and --tts all win. Use it with serve instead of local when you want to expose the server without starting the microphone/speaker client.

--tts pocket and --tts kokoro are also valid on macOS.

To compare the MLX quantization variants locally:

python scripts/benchmark_tts.py \
    --handlers qwen3 \
    --iterations 3 \
    --qwen3_mlx_quantizations bf16 4bit 6bit 8bit

Install the NVIDIA Container Toolkit, then:

docker compose up

The compose file starts a llama.cpp server with Gemma 4 and the Realtime server, exposing ports 8080 and 8765.

Realtime mode supports the OpenAI Realtime protocol over WebSocket and WebRTC, with live transcription and low-latency turn-taking. WebSocket clients connect at /v1/realtime:

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8765/v1",
    websocket_base_url="ws://localhost:8765/v1",
    api_key="not-needed",
)

with client.realtime.connect(model="local") as conn:
    conn.send(
        {
            "type": "session.update",
            "session": {
                "type": "realtime",
                "instructions": "You are a helpful assistant.",
                "audio": {
                    "input": {
                        "turn_detection": {
                            "type": "server_vad",
                            "interrupt_response": True,
                        }
                    }
                },
            },
        }
    )

    for event in conn:
        print(event.type)

The server implements the core Realtime event set: input_audio_buffer.append, session.update, conversation.item.create, response.create, and response.cancel inbound; speech start/stop, streaming transcription, audio deltas, tool calls, and response.done outbound. The full event reference, architecture, and design details live in the Realtime Engine README.

With --enable_llm_proxy, the realtime server also exposes the remote LLM it is configured with as a plain OpenAI compatible endpoint, so a client can run side tasks (summaries, titles, background agents) with tools and streaming, fully concurrent with the voice conversation and never interrupted by new speech:

• POST /v1/chat/completions when running --llm_backend chat-completions
• POST /v1/responses when running --llm_backend responses-api

The server performs no authentication and no throttling of its own. Enable the proxy only on a trusted network, or deploy the server behind a gateway that owns access control. The s2s-endpoint compute replica is such a gateway: it opens these paths only to clients that created their session with an HF token, checks the API key against that token, and applies a rate limit per user. Point the stock OpenAI SDK at whichever host you talk to; this server ignores the API key (a gateway in front decides what it must be):

from openai import OpenAI

llm = OpenAI(base_url="http://localhost:8765/v1", api_key="unused")
completion = llm.chat.completions.create(
    model="anything",  # ignored: the server forces its configured --model_name
    messages=[{"role": "user", "content": "Summarize the conversation so far: ..."}],
)

Requests are stateless (send the full message list each time) and are proxied to the configured upstream with the key held by the server, which never reaches clients. The model field is always overwritten with the server configured --model_name. The proxy is off by default, requires a remote backend (chat-completions or responses-api), and answers 501 with the reason otherwise.

The LLM is the most compute-intensive and highest-latency component in the pipeline. A single forward pass through a large model can dominate end-to-end response time, so choosing the right backend for your hardware and latency budget matters. The pipeline supports:

• Local inference: transformers on CUDA / CPU and mlx-lm on Apple Silicon.
• Self-hosted servers: responses-api and chat-completions can point at a local vLLM or llama.cpp server.
• Provider APIs: the same backends work with OpenAI, HF Inference Providers, OpenRouter, and other OpenAI-compatible providers.

Two API backends are available, sharing the same --responses_api_* connection flags:

• --llm_backend responses-api (default) targets /v1/responses.
• --llm_backend chat-completions targets /v1/chat/completions.

Use --stt none --llm_backend chat-completions to send each completed VAD audio segment directly to an audio-input model. Direct audio mode is not supported with --llm_backend responses-api: a model may accept audio through /v1/chat/completions without supporting /v1/responses, including OpenAI's gpt-audio-1.5.

You must explicitly set --model_name to a model that accepts audio: the default gpt-5.4-mini accepts text and image input, but not audio. Check the provider's model documentation and endpoint support before enabling this mode. For OpenAI, see the GPT-5.4 mini model card and audio-input guide.

speech-to-speech serve \
    --stt none \
    --llm_backend chat-completions \
    --model_name "YOUR_AUDIO_CAPABLE_MODEL" \
    --responses_api_base_url "https://provider.example/v1" \
    --responses_api_api_key "$PROVIDER_API_KEY"

OpenAI-compatible servers represent input audio differently. Use --responses_api_audio_content_type input_audio (the default) for embedded WAV base64, or --responses_api_audio_content_type audio_url for a base64 data URL.

The examples below pair Parakeet TDT for local STT and Qwen3-TTS for local TTS with different LLM backends.

Works with any provider or server that implements the OpenAI Responses API. Point --responses_api_base_url at the endpoint and set --model_name accordingly:

# OpenAI
speech-to-speech local \
    --stt parakeet-tdt \
    --llm_backend responses-api \
    --tts qwen3 \
    --qwen3_tts_mlx_quantization 6bit \
    --model_name "gpt-4o-mini" \
    --responses_api_api_key "$OPENAI_API_KEY" \
    --responses_api_stream \
    --enable_live_transcription

# HF Inference Providers: Qwen3.5-9B via Together
speech-to-speech local \
    --stt parakeet-tdt \
    --llm_backend responses-api \
    --tts qwen3 \
    --qwen3_tts_mlx_quantization 6bit \
    --model_name "Qwen/Qwen3.5-9B:together" \
    --responses_api_base_url "https://router.huggingface.co/v1" \
    --responses_api_api_key "$HF_TOKEN" \
    --responses_api_stream \
    --enable_live_transcription

# HF Inference Providers: GPT-oss-20B via Groq
speech-to-speech serve \
    --stt parakeet-tdt \
    --llm_backend responses-api \
    --tts qwen3 \
    --qwen3_tts_mlx_quantization 6bit \
    --model_name "openai/gpt-oss-20b:groq" \
    --responses_api_base_url "https://router.huggingface.co/v1" \
    --responses_api_api_key "$HF_TOKEN" \
    --responses_api_stream \
    --enable_live_transcription

Identical configuration to responses-api, reusing the same --responses_api_* connection flags, but talks to /v1/chat/completions instead of /v1/responses. Prefer it when:

• the provider ignores chat_template_kwargs.enable_thinking on the Responses path and needs a reasoning_effort knob to suppress reasoning, or
• the server's Responses streaming tool-call path is unreliable, while its Chat Completions tool-call streaming is solid. This is useful for some vLLM builds; see #312.

Add --responses_api_reasoning_effort none to disable reasoning on providers where the chat-template flag has no effect:

# vLLM serving a Qwen model with tool calling
speech-to-speech serve \
    --stt parakeet-tdt \
    --llm_backend chat-completions \
    --tts qwen3 \
    --model_name "Qwen/Qwen3-4B-Instruct-2507" \
    --responses_api_base_url "http://localhost:8000/v1" \
    --responses_api_stream

# Gemma 4 31B via the HF router on Cerebras, with reasoning disabled for low voice latency
speech-to-speech serve \
    --stt parakeet-tdt \
    --llm_backend chat-completions \
    --tts qwen3 \
    --model_name "google/gemma-4-31B-it:cerebras" \
    --responses_api_base_url "https://router.huggingface.co/v1" \
    --responses_api_api_key "$HF_TOKEN" \
    --responses_api_reasoning_effort none \
    --responses_api_stream

Run the LLM in a separate llama.cpp process for the lowest-friction fully local setup, as shown in the Reachy Mini local conversation guide:

For a fully local native-audio setup with the browser demo, Realtime turn revisions, and barge-in, see the tested Gemma 4 12B speech-to-speech example for Apple Silicon.

# Terminal 1: llama.cpp serving Gemma 4
llama-server -hf ggml-org/gemma-4-E4B-it-GGUF -np 2 -c 65536 -fa on --swa-full

# Terminal 2: speech-to-speech using that local LLM server
speech-to-speech serve \
    --stt parakeet-tdt \
    --llm_backend responses-api \
    --tts qwen3 \
    --model_name "ggml-org/gemma-4-E4B-it-GGUF" \
    --responses_api_base_url "http://127.0.0.1:8080/v1" \
    --responses_api_api_key "" \
    --responses_api_stream \
    --enable_live_transcription

Use speech-to-speech local when you want to run the same server and talk through the machine hosting it. In-process local backends are available with --llm_backend mlx-lm on Apple Silicon or --llm_backend transformers on CUDA / CPU.

Language coverage depends on the STT and TTS backends you pick, not on the pipeline itself:

Make sure the STT, LLM, and TTS you pair all cover your target language(s). Two usage patterns:

• Single language: set --language to the target language code. The default is en.
• Language switching: set --language auto. The STT detects the language of each spoken prompt and forwards it to the LLM. Optionally add --enable_lang_prompt to append a "Please reply to my message in ..." instruction. It defaults to False; large LLMs usually infer the language from context, but the explicit instruction can help smaller models.

Automatic language detection:

speech-to-speech serve \
    --stt parakeet-tdt \
    --language auto \
    --llm_backend mlx-lm \
    --model_name "mlx-community/Qwen3-4B-Instruct-2507-bf16"

A single non-English language, Chinese in this example:

speech-to-speech serve \
    --stt whisper-mlx \
    --stt_model_name large-v3 \
    --language zh \
    --llm_backend mlx-lm \
    --model_name mlx-community/Qwen3-4B-Instruct-2507-bf16

Both commands also work with --mac-optimal-settings; explicit --stt flags override the defaults it sets.

Pocket TTS from Kyutai Labs provides streaming TTS with voice cloning:

speech-to-speech serve \
    --tts pocket \
    --pocket_tts_voice jean \
    --pocket_tts_device cpu

Available voice presets: alba, marius, javert, jean, fantine, cosette, eponine, azelma. Custom voice files and Hugging Face paths also work.

References for pipeline CLI arguments live in the arguments classes and in speech-to-speech serve -h. Client arguments are listed by speech-to-speech talk -h.

See ModuleArguments. It allows setting:

• a common --device, if every part should run on the same device
• macOS model/device defaults (--mac-optimal-settings)
• STT implementation (--stt)
• LLM backend (--llm_backend: transformers, mlx-lm, responses-api, or chat-completions)
• TTS implementation (--tts)
• logging level
• realtime pipeline pool size (--num_pipelines)

See VADHandlerArguments. Notable options:

• --thresh: threshold value to trigger voice activity detection.
• --min_speech_ms: minimum duration of detected voice activity to be considered speech.
• --min_speech_continuation_ms: sustain-bar hysteresis threshold for speech that continues a reopenable soft-ended, uncommitted turn within the reopen window. The default and recommended pairing is --min_speech_ms 384 --min_speech_continuation_ms 192.
• --min_silence_ms: minimum length of silence intervals for segmenting speech. Default is 64 ms.
• --short_segment_merge_ms: optional merge window for stitching adjacent VAD segments that are each shorter than --min_speech_ms.
• --speculative_reopen_ms: delay response commitment for 800 ms after a soft-ended turn so immediately resumed speech can reopen it.
• --unanswered_reopen_ms: sanity cap on how long a soft-ended speculative turn that has not yet received any assistant output stays reopenable. With Smart Turn enabled, this is clamped to at least --smart_turn_max_wait_ms so a turn remains reopenable for its full grace.

Smart Turn v3.2 can validate Silero's end-of-speech decisions using the content and prosody of the current turn. Silero finalizes the segment and STT/LLM work may begin speculatively. Complete turns start processing immediately and use --speculative_reopen_ms (800 ms by default) before committing output. Incomplete turns wait --smart_turn_incomplete_delay_ms (600 ms by default) before starting STT/LLM work, while their output remains gated by --smart_turn_max_wait_ms (2 seconds by default). If speech resumes during either delay, the existing turn is reopened as a newer revision, the accumulated audio is re-emitted, and work from the previous revision is discarded before it reaches the user.

The base package includes the quantized CPU runtime and enables Smart Turn by default:

pip install speech-to-speech
speech-to-speech serve

The latest supported v3.2 CPU checkpoint downloads from the Hugging Face Hub on first use. Pass --smart_turn_model_path /path/to/model.onnx to use a local model, or --no_smart_turn to disable Smart Turn. Smart Turn is enabled by default for server sessions and the packaged local client.

Tune the completion cutoff with --smart_turn_threshold (default 0.5). A higher threshold makes ambiguous pauses more likely to use the longer speculative response grace.

model_name, torch_dtype, and device are exposed for each STT, LLM, and TTS implementation. STT and TTS parameters use the handler prefix, for example --stt_model_name or --qwen3_tts_device. LLM model selection and chat settings are shared across backends via unprefixed flags, for example --model_name and --chat_size; backend-specific flags use the responses_api_ prefix for the responses-api and chat-completions backends and the llm_ prefix for local backends.

For example:

# Local transformers/mlx-lm backend
--model_name google/gemma-2b-it

# OpenAI-compatible backend
--llm_backend responses-api --model_name deepseek-chat --responses_api_base_url https://api.deepseek.com

Other generation parameters can be set using the handler prefix plus _gen_, for example --stt_gen_max_new_tokens 128 or --llm_gen_temperature 0.7. Parameters not yet exposed can be added to the relevant arguments class.

Issues and PRs are welcome. Good starting points are the open issues. For larger changes, open an issue first to discuss the approach.

For local development:

uv sync
pytest
ruff check

If you use this pipeline, please also cite the component models you run. The defaults are:

@misc{SileroVAD,
  author = {Silero Team},
  title = {Silero VAD: pre-trained enterprise-grade Voice Activity Detector (VAD), Number Detector and Language Classifier},
  year = {2021},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/snakers4/silero-vad}},
  email = {hello@silero.ai}
}

@misc{parakeet-tdt,
  author = {NVIDIA},
  title = {Parakeet TDT 0.6B v3},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3}}
}

@misc{qwen3-tts,
  author = {Qwen Team},
  title = {Qwen3-TTS},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice}}
}

Citations for optional backends such as Kokoro, Pocket TTS, ChatTTS, Whisper variants, Paraformer, and MMS live in the respective component READMEs.