Text to Speech

Get started

calibrate-agent tts

The interactive UI guides you through the full evaluation process:

Language selection — pick from 10+ supported Indic languages
Provider selection — choose providers (only those supporting your language are shown)
Input CSV — path to CSV file with id and text columns

The input CSV should have this format:

id	text
row_1	hello world
row_2	this is a test
row_3	how are you doing today

Refer to this sample for a template.

Output directory — where results will be saved (defaults to ./out)
API keys — enter the API keys for the selected providers

The evaluation runs providers in parallel (max 2 at a time), showing the progress as audio files are generated.

Evaluator configuration

By default, an audio LLM judge — routed through OpenRouter (set OPENROUTER_API_KEY in your environment) — evaluates whether the reference text is pronounced correctly in the synthesized audio using the built-in pronunciation evaluator; expand Default evaluator: pronunciation below for the exact system_prompt from the codebase. You can customize the judge model and add multiple evaluators by passing an optional config file with --config:

calibrate-agent tts -p openai google -i sample.csv -o ./out --config config.json

Each evaluator’s system_prompt is sent as the system message to its own dedicated audio LLM judge call (one call per evaluator, run in parallel). The user message contains the reference text and the audio sample. The config file supports:

{
  "evaluators": [
    {
      "id": "intelligibility-id",
      "name": "intelligibility",
      "system_prompt": "You are a highly accurate evaluator. You will be given an audio sample and the reference text it is supposed to speak. Mark True if the spoken text is clearly understandable from the audio.",
      "judge_model": "openai/gpt-audio"
    },
    {
      "id": "pronunciation-id",
      "name": "pronunciation",
      "system_prompt": "You are a highly accurate evaluator. You will be given an audio sample and the reference text it is supposed to speak. Mark True only if all words are pronounced correctly with natural-sounding speech.",
      "judge_model": "openai/gpt-audio"
    }
  ]
}

Key	Type	Description
`evaluators`	array	List of evaluators. Each one becomes its own audio LLM call per row.
`evaluators[].id`	string	Optional unique id. Output `config.json` includes the raw `evaluators` list and an `evaluators_map` from id to name.
`evaluators[].name`	string	Unique evaluator name. Becomes the column name in the leaderboard.
`evaluators[].system_prompt`	string	Full system prompt used for this evaluator’s audio LLM judge call.
`evaluators[].judge_model`	string	OpenRouter model id for this evaluator (default: `openai/gpt-audio`). Must be an audio-capable model in the OpenRouter catalog — for example OpenAI’s audio models or Google’s Gemini audio-capable entries; the sample config uses `google/gemini-2.5-flash`.

Each evaluator also accepts:

Key	Type	Description
`type`	string	`"binary"` (default) or `"rating"`
`scale_min`	integer	Required when `type` is `"rating"`. Lowest allowed score.
`scale_max`	integer	Required when `type` is `"rating"`. Highest allowed score.

Binary evaluators produce per-row pass/fail and a mean pass-rate. Rating evaluators produce an integer score on your scale and a mean score in the leaderboard. When multiple evaluators are defined, each is scored independently — one audio LLM call per evaluator per row, all run in parallel — and appears as a separate column in the results and leaderboard. Refer to the sample config for a template.

The --config flag is optional. When omitted, a single built-in pronunciation evaluator scores audio intelligibility. The TTS judge requires an audio-capable model.

pronunciation (default evaluator system prompt)

Matches DEFAULT_TTS_EVALUATOR in calibrate/judges.py when no --config is passed.

You are a highly accurate evaluator evaluating the audio output of a TTS model.

You will be given the audio and the text that should have been spoken in the audio.

You need to evaluate if the text is easily understandable from the audio. Check whether the spoken words match the reference text and the audio is clear enough to convey the intended message.

Output

Once all the providers have completed, it displays a leaderboard measuring key metrics along with bar charts for better visualization. Each provider output directory includes:

results.csv with one row per input text (id, text, audio_path, ttfb, evaluator scores, and evaluator reasoning).
metrics.json with evaluator summaries, ttfb percentiles, and a cost object. Each provider is priced in its native billing unit — no unit conversion. Character-billed providers (most) report billing_unit: "character", total_characters, and the per-million-character rate (cost_per_million_chars_currency); audio-billed providers (OpenAI, Gemini) report billing_unit: "minute", total_seconds, audio_minutes, and the per-minute rate (cost_per_minute_currency), with the audio duration measured from the synthesized output. The rate is denominated in currency. Providers billed in a non-USD currency (e.g. Sarvam in INR) report the native-currency total cost_in_currency and, when the live exchange rate is reachable, the conversion_rate used (native-currency units per 1 USD) and cost_usd. If the exchange rate can’t be fetched, the native-currency cost is still reported and cost_usd is omitted, so that provider is left out of the USD comparison rather than failing the run. cost_usd is the metric comparable across all providers.
leaderboard/tts_leaderboard.xlsx with a cost_usd column in the summary sheet (the one cost figure comparable across providers; per-unit rates and native-currency figures stay in each provider’s metrics.json).

You can also view the generated audio and metrics for each row of your dataset including the LLM judge score and reasoning. Use the arrow keys to navigate rows and press Enter or p to play the generated audio.

Getting started

Guides

Get started

Evaluator configuration

Output

Learn more about metrics

Resources

Integrations

​Get started

​Evaluator configuration

​Output

Learn more about metrics

​Resources

Integrations

Get started

Evaluator configuration

Output

Resources