enne2/qwentts.cpp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

server: sampling overrides on the speech endpoint

Browse Source 
The speech body accepts seed, max_new_tokens, temperature, top_k,
top_p, and repetition_penalty. Unset fields keep the engine defaults,
a temperature of zero selects greedy decoding, and the subtalker
mirrors the talker knobs. A fixed seed makes a request reproducible.
This commit is contained in:
Pascal
2026-07-05 15:01:37 +02:00
parent 1a680e8816
commit 73fe0c67bb
4 changed files with 100 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+7 -1
View File
@@ -139,9 +139,15 @@ curl -X POST localhost:8080/v1/voices -H "Content-Type: application/json" \
\"spk_b64\":\"$(base64 -w0 ref.spk)\",\"rvq_b64\":\"$(base64 -w0 ref.rvq)\"}"
curl -X POST localhost:8080/v1/audio/speech -H "Content-Type: application/json" \
-d '{"input":"Hello world.","voice":"freeman","response_format":"wav"}' -o out.wav
-d '{"input":"Hello world.","voice":"freeman","response_format":"wav",
"seed":42,"temperature":0.8}' -o out.wav
```
The speech body accepts optional sampling overrides (`seed`,
`max_new_tokens`, `temperature`, `top_k`, `top_p`,
`repetition_penalty`); unset fields keep the engine defaults and a
fixed seed makes the request reproducible.
## Embedding the library
The CLI tools are thin wrappers over a public ABI. Single-header,
docs/ARCHITECTURE.md
+7 -1
View File
@@ -626,7 +626,13 @@ Endpoints :
```
POST /v1/audio/speech OAI text-to-speech; response_format "pcm"
streams s16le 24 kHz mono chunked as it is
generated, "wav" returns a one-shot RIFF file
generated, "wav" returns a one-shot RIFF file.
Optional sampling overrides ride in the same
body: seed, max_new_tokens, temperature,
top_k, top_p, repetition_penalty. Unset
fields keep the engine defaults, temperature
0 selects greedy decoding, the subtalker
mirrors the talker knobs
GET /v1/models single loaded model
GET /v1/voices model speakers plus registered cloned voices
POST /v1/voices register a cloned voice: {name, ref_text,
src/tts-server.h
+56
View File
@@ -26,6 +26,7 @@
#include "audio-io.h"
#include "yyjson.h"
#include <cfloat>
#include <cmath>
#include <csignal>
#include <cstdint>
@@ -42,6 +43,15 @@ struct tts_request {
std::string instructions; // OAI instructions, mapped to the ABI instruct field
std::string format; // "pcm" (stream) or "wav" (one-shot)
float speed; // OAI speed, parsed then ignored (no time stretch in the ABI)
// Optional sampling overrides. -1 (ints) and NaN (floats) mark a
// field the client left unset, keeping the engine defaults.
int64_t seed; // forwarded verbatim, -1 draws a random seed
int max_new_tokens; // strictly positive
int top_k; // 0 disables the top-k filter
float temperature; // 0 selects greedy decoding
float top_p; // in (0, 1]
float repetition_penalty; // strictly positive
};
// One voice registration parsed from the POST /v1/voices JSON body.
@@ -165,8 +175,54 @@ static bool tts_parse_request(const std::string & body, tts_request & req, std::
yyjson_val * speed = yyjson_obj_get(root, "speed");
req.speed = yyjson_is_num(speed) ? (float) yyjson_get_num(speed) : 1.0f;
// Optional sampling overrides. A missing field keeps its unset
// marker; a present field must be well typed and in domain.
req.seed = -1;
req.max_new_tokens = -1;
req.top_k = -1;
req.temperature = NAN;
req.top_p = NAN;
req.repetition_penalty = NAN;
auto opt_int = [&](const char * key, int64_t lo, int64_t hi, int64_t & out) -> bool {
yyjson_val * v = yyjson_obj_get(root, key);
if (!v) {
return true;
}
if (!yyjson_is_int(v) || yyjson_get_sint(v) < lo || yyjson_get_sint(v) > hi) {
err = std::string("'") + key + "' is out of domain";
return false;
}
out = yyjson_get_sint(v);
return true;
};
auto opt_num = [&](const char * key, double lo, double hi, float & out) -> bool {
yyjson_val * v = yyjson_obj_get(root, key);
if (!v) {
return true;
}
if (!yyjson_is_num(v) || yyjson_get_num(v) < lo || yyjson_get_num(v) > hi) {
err = std::string("'") + key + "' is out of domain";
return false;
}
out = (float) yyjson_get_num(v);
return true;
};
int64_t max_new = -1;
int64_t top_k = -1;
bool ok = opt_int("seed", INT64_MIN, INT64_MAX, req.seed) && opt_int("max_new_tokens", 1, INT32_MAX, max_new) &&
opt_int("top_k", 0, INT32_MAX, top_k) && opt_num("temperature", 0.0, FLT_MAX, req.temperature) &&
opt_num("top_p", DBL_MIN, 1.0, req.top_p) &&
opt_num("repetition_penalty", DBL_MIN, FLT_MAX, req.repetition_penalty);
req.max_new_tokens = (int) max_new;
req.top_k = (int) top_k;
yyjson_doc_free(doc);
if (!ok) {
return false;
}
if (req.format != "pcm" && req.format != "wav") {
err = "response_format must be 'pcm' or 'wav'";
return false;
tools/tts-server.cpp
+30
View File
@@ -9,6 +9,7 @@
#include "rvq-file.h"
#include "version.h"
#include <cmath>
#include <cstdio>
#include <cstdlib>
#include <cstring>
@@ -227,6 +228,35 @@ int main(int argc, char ** argv) {
p.instruct = req.instructions.c_str();
}
// Sampling overrides ride straight into the ABI; the subtalker
// mirrors the talker knobs so the HTTP surface stays a single
// coherent set. A temperature of zero selects greedy decoding
// on both.
p.seed = req.seed;
if (req.max_new_tokens != -1) {
p.max_new_tokens = req.max_new_tokens;
}
if (req.top_k != -1) {
p.top_k = req.top_k;
p.subtalker_top_k = req.top_k;
}
if (!std::isnan(req.temperature)) {
if (req.temperature == 0.0f) {
p.do_sample = false;
p.subtalker_do_sample = false;
} else {
p.temperature = req.temperature;
p.subtalker_temperature = req.temperature;
}
}
if (!std::isnan(req.top_p)) {
p.top_p = req.top_p;
p.subtalker_top_p = req.top_p;
}
if (!std::isnan(req.repetition_penalty)) {
p.repetition_penalty = req.repetition_penalty;
}
// Trampoline : the C ABI on_chunk forwards to the C++ sink.
const tts_sink * sink_ptr = &sink;
p.on_chunk = [](const float * s, int ns, void * u) -> bool {