Grafy News
Try Grafy free

Bland Introduces Tofu-1, a Full‑Duplex Speech‑to‑Speech Model for Enterprise Calls

Bland has released Tofu-1, a full‑duplex speech‑to‑speech model aimed at noisy, high‑stakes phone calls, now in enterprise research preview.

Illustration for the news story: Bland Introduces Tofu-1, a Full‑Duplex Speech‑to‑Speech Model for Enterprise CallsSource: MarkTechPost

Bland has rolled out Tofu-1, a full‑duplex speech‑to‑speech model designed for noisy, high‑stakes phone calls, and the release is being offered as an enterprise research preview. The announcement matters because it promises real‑time, two‑way voice conversion that could reshape how businesses handle customer support, remote conferencing, and other mission‑critical communications.

What is Tofu-1 and why does it matter?

According to MarkTechPost, Tofu-1 is a speech‑to‑speech system that processes both inbound and outbound audio streams simultaneously, a capability known as full‑duplex. Traditional speech‑to‑text or text‑to‑speech pipelines handle one direction at a time, introducing latency and requiring separate components for each side of a conversation. By handling both directions in a single model, Tofu-1 aims to reduce delay, preserve conversational flow, and maintain intelligibility even in environments with substantial background noise.

How does full‑duplex speech‑to‑speech differ from existing solutions?

Most commercial voice AI products rely on a cascade of separate models: a speech‑to‑text engine, a language understanding module, and a text‑to‑speech synthesizer. Each stage adds processing time and potential error propagation. Full‑duplex speech‑to‑speech models, such as Tofu-1, skip the intermediate text representation, directly mapping a source language’s acoustic features to target acoustic features while simultaneously listening and speaking. This architecture can:

  • Cut round‑trip latency, which is critical for real‑time negotiations or emergency response calls.
  • Better handle overlapping speech, because the model keeps a continuous representation of both streams.
  • Offer more consistent audio quality when background noise varies, as the model is trained on noisy call data.

While the press release does not provide benchmark numbers, the claim of handling “noisy, high‑stakes” calls suggests a training set that includes real‑world telephone audio rather than clean studio recordings.

What are the limitations of the current preview?

The MarkTechPost article notes that Tofu-1 is still in an enterprise research preview. At this stage, Bland has not published:

  • Model size (parameter count or memory footprint).
  • Quantitative performance metrics (e.g., word error rate, signal‑to‑noise ratio improvement).
  • Pricing or licensing details.

Without these data points, developers cannot accurately gauge resource requirements or cost‑effectiveness. Additionally, the preview is limited to enterprise partners, meaning independent developers or hobbyists cannot yet access the model directly.

Potential use cases for enterprises

Even with limited public information, several practical scenarios emerge:

  1. Customer support centers – Agents could receive a clean, amplified version of a caller’s voice while the system simultaneously generates a high‑quality response in the agent’s preferred language or accent.
  2. Financial trading desks – Rapid, error‑free voice communication is essential during market volatility; a full‑duplex model could reduce miscommunication caused by noisy floor environments.
  3. Healthcare tele‑triage – Doctors often need to understand patients over poor‑quality lines; Tofu-1 could improve intelligibility without requiring additional hardware.
  4. Multilingual conference calls – By pairing Tofu-1 with a language‑translation layer, enterprises could run real‑time, two‑way translated meetings without separate transcription and synthesis pipelines.

How to experiment with Tofu‑1 using Grafy

While Grafy does not yet host Tofu-1 as a built‑in model, creators and researchers can still explore the concept through related tools:

  • Grafy Files – Upload recorded call audio files (e.g., .wav or .mp3) and use the built‑in noise‑reduction and transcription utilities to simulate the inbound processing stage of a full‑duplex system. The Files app’s audio conversion features let you experiment with different sample rates and channel configurations.
  • Grafy Chat – Connect to external APIs that expose Tofu-1 (once Bland opens access) by selecting a custom endpoint in the Chat interface. This lets developers prototype prompt‑style interactions, feeding raw audio chunks and receiving synthesized responses in real time.
  • Grafy Graph – Document the workflow of a full‑duplex pipeline, branching each step (input capture, model inference, output playback) and visualising performance metrics as they become available.

To get started, you could:

  1. Record a short noisy phone conversation.
  2. Upload the file to /files and apply noise suppression.
  3. Use /chat to call a placeholder speech‑to‑speech endpoint (e.g., a public Whisper or Tacotron API) to mimic one half of the duplex flow.
  4. Capture the output and re‑inject it as the next input, iterating to approximate a two‑way exchange.

When Bland releases an official API key, swapping the placeholder endpoint for Tofu‑1 will turn this prototype into a live test.

Frequently asked questions

Does Tofu-1 support languages other than English?

The preview announcement does not list supported languages. Historically, full‑duplex speech‑to‑speech research focuses first on English, but enterprise customers often request multilingual capability, so future releases may include additional languages.

Can developers use Tofu-1 for personal projects or open‑source research?

At the moment, Bland limits access to an enterprise research preview, meaning only partnered businesses can evaluate the model. Independent developers will need to wait for a public release or a limited‑access program.

How does Tofu-1 handle overlapping speech from both parties?

Full‑duplex architectures are designed to process simultaneous audio streams, which helps the model separate and reconstruct each speaker’s voice even when they talk over each other. Specific performance figures have not been disclosed yet.


Source: MarkTechPost — Bland Launches Tofu-1, a Full-Duplex Speech-to-Speech Model for Calls