Patrick Desjardins Blog

Posted on: 2026-03-25

I recently had many voices to generate from English text. By many, I mean about 5,000 paragraphs. One requirement I had was to produce a non-robotic voice. After searching, I was open to using a service like OpenAI TTS and paying for it, but it was quite a large amount of text to generate. Since I have a good graphics card (Nvidia 5080), I preferred using it locally. I found: Kokoro TTS.

The TTS relies on a model with about 82 million parameters, which is relatively lightweight and can run locally. It can also accept large amounts of text and provides over a dozen English voices.

Kokoro has a CLI but also a Python package that allows several ways to generate audio files. The model is automatically downloaded and installed the first time the Kokoro pipeline is instantiated.

Here is roughly what is needed:

try:
    pipeline = KPipeline(lang_code="a")
    clips = []
    phonemes = []
    for _, phoneme_string, audio in pipeline(args.text, voice=args.voice, speed=args.speed):
        if phoneme_string:
            phonemes.append(str(phoneme_string))
        if audio is not None:
            clips.append(np.asarray(audio, dtype=np.float32))

    if not clips:
        print("Kokoro returned no audio for this text.", file=sys.stderr)
        return 1

    merged = np.concatenate(clips)
    sf.write(args.out, merged, 24000)
    print(json.dumps({"out": args.out, "voice": args.voice, "phonemes": " ".join(phonemes)}))
    return 0
except Exception as error:
    print(f"Kokoro generation failed: {error}", file=sys.stderr)
    return 1

The loop works with a generator that returns audio segments progressively as the text is synthesized. The output contains audio segments that are merged and written into a sound file. The pipeline also returns phonemes, which I am not using yet but could be useful in the future to explain pronunciation through another AI tool.

The generated audio uses a 24 kHz sample rate, which is the native output of the Kokoro models. The final merged waveform is written directly to a file using this sample rate. On my Nvidia 5080 GPU, generation is fast enough to process thousands of paragraphs in a reasonable time (few minutes). The lightweight 82M parameter model keeps memory usage low and allows batch processing without the overhead typically associated with larger speech models.

So far, the results have been significantly better than the edge-tts Python library while remaining completely free.